Robotics: Science and Systems VI 0262516810, 9780262516815

Citation preview

Robotics

Robotics Science and Systems VI

edited by Yoky Matsuoka, Hugh Durrant-Whyte, and Jos´e Neira

The MIT Press Cambridge, Massachusetts London, England

c 2011 Massachusetts Institute of Technology  All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means (including photocopying, recording, or information storage and retrieval) without permission in writing from the publisher. For information about special quantity discounts, please email special [email protected] or write to Special Sales Department, The MIT Press, 55 Hayward Street, Cambridge, MA 02142. Printed and bound in the United States of America. Library of Congress Cataloging-in-Publication Data Robotics: Science and Systems Conference (6th : 2010 : Zaragoza, Spain) Robotics : science and systems VI / edited by Yoky Matsuoka, Hugh Durrant-Whyte, and Jos´e Neira. p. cm. “This volume contains the 40 papers presented at Robotics: Science and Systems (RSS) 2010, held at the University of Zaragoza in Spain, from June 27 to June 30, 2010”—Pref. Includes bibliographical references. ISBN 978-0-262-51681-5 (pbk. : alk. paper) 1. Robotics—Congresses. I. Matsuoka, Yoky. II. Durrant-Whyte, Hugh F., 1961- III. Neira, Jos´e. VI. Title. TJ210.3.R6435 2011 629.8’92—dc22 2011007422

10 9 8 7 6 5 4 3 2 1

Contents Preface

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ix

Organizing Committee . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii Program Committee Sponsors

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xv

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii

Biophysically Inspired Development of a Sand-Swimming Robot Ryan D. Maladen, Yang Ding, Paul B. Umbanhowar, Adam Kamor, and Daniel I. Goldman .

1

Passive Torque Regulation in an Underactuated Flapping Wing Robotic Insect P. S. Sreetharan and R. J. Wood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

9

Color-Accurate Underwater Imaging Using Perceptual Adaptive Illumination Iuliu Vasilescu, Carrick Detweiler, and Daniela Rus . . . . . . . . . . . . . . . . . . . . .

17

Probabilistic Lane Estimation Using Basis Curves Albert S. Huang and Seth Teller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

25

Reinforcement Learning to Adjust Robot Movements to New Situations Jens Kober, Erhan Oztop, and Jan Peters . . . . . . . . . . . . . . . . . . . . . . . . . . .

33

Analysis and Control of a Dissipative Spring-Mass Hopper with Torque Actuation M. Mert Ankaralı and Uluc¸ Saranlı . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

41

On Motion and Force Control of Grasping Hands with Postural Synergies D. Prattichizzo, M. Malvezzi, and A. Bicchi . . . . . . . . . . . . . . . . . . . . . . . . . .

49

Back-drivable and Inherently Safe Mechanism for Artificial Finger Koichi Koganezawa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

57

Segmentation and Unsupervised Part-based Discovery of Repetitive Objects Rudolph Triebel, Jiwon Shin, and Roland Siegwart . . . . . . . . . . . . . . . . . . . . . .

65

Scale Drift-Aware Large Scale Monocular SLAM Hauke Strasdat, J. M. M. Montiel, and Andrew J. Davison

. . . . . . . . . . . . . . . . . .

73

Preliminary Results in Decentralized Estimation for Single-Beacon Acoustic Underwater Navigation Sarah E. Webster, Louis L. Whitcomb, and Ryan M. Eustice . . . . . . . . . . . . . . . . . .

81

A Non-invasive, Real-Time Method for Measuring Variable Stiffness Giorgio Grioli and Antonio Bicchi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

89

Consistent Data Association in Multi-robot Systems with Limited Communications Rosario Arag¨ue´ s, Eduardo Montijano, and Carlos Sag¨ue´ s . . . . . . . . . . . . . . . . . .

97



Singularity-invariant Leg Rearrangements in Doubly-planar Stewart-Gough Platforms J´ulia Borr`as, Federico Thomas, and Carme Torras . . . . . . . . . . . . . . . . . . . . . . 105 On the Kinematic Design of Exoskeletons and Their Fixations with a Human Member Nathana¨el Jarrass´e and Guillaume Morel . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 Assessing Optimal Assignment under Uncertainty: An Interval-based Algorithm Lantao Liu and Dylan A. Shell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 LQG-MP: Optimized Path Planning for Robots with Motion Uncertainty and Imperfect State Information Jur van den Berg, Pieter Abbeel, and Ken Goldberg . . . . . . . . . . . . . . . . . . . . . . 129 The Smooth Curvature Flexure Model: An Accurate, Low-dimensional Approach for Robot Analysis Lael U. Odhner and Aaron M. Dollar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 Multi-priority Cartesian Impedance Control Robert Platt Jr., Muhammad Abdallah, and Charles Wampler

. . . . . . . . . . . . . . . . 145

Variable Impedance Control: A Reinforcement Learning Approach Jonas Buchli, Evangelos Theodorou, Freek Stulp, and Stefan Schaal . . . . . . . . . . . . . 153 A Fast Traversal Heuristic and Optimal Algorithm for Effective Environmental Coverage Ling Xu and Tony Stentz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 Stochastic Complementarity for Local Control of Discontinuous Dynamics Yuval Tassa and Emo Todorov . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 Distributed Optimization with Pairwise Constraints and Its Application to Multi-robot Path Planning Subhrajit Bhattacharya, Vijay Kumar, and Maxim Likhachev . . . . . . . . . . . . . . . . . 177 PLISS: Detecting and Labeling Places Using Online Change-Point Detection Ananth Ranganathan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 A Constant-Time Algorithm for Vector Field SLAM Using an Exactly Sparse Extended Information Filter Jens-Steffen Gutmann, Ethan Eade, Philip Fong, and Mario Munich . . . . . . . . . . . . . 193 Efficient Probabilistic Planar Robot Motion Estimation Given Pairs of Images Olaf Booij, Ben Kr¨ose, and Zoran Zivkovic . . . . . . . . . . . . . . . . . . . . . . . . . . 201 Efficient Non-parametric Surface Representations Using Active Sampling for Push Broom Laser Data Mike Smith, Ingmar Posner, and Paul Newman . . . . . . . . . . . . . . . . . . . . . . . . 209 Sensor Placement for Improved Robotic Navigation Michael P. Vitus and Claire J. Tomlin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 Task-driven Tactile Exploration Kaijen Hsiao, Leslie Pack Kaelbling, and Tom´as Lozano-P´erez . . . . . . . . . . . . . . . . 225 On the Role of Hand Synergies in the Optimal Choice of Grasping Forces Marco Gabiccini and Antonio Bicchi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233  vi

Dynamic Constraint-based Optimal Shape Trajectory Planner for Shape-Accelerated Underactuated Balancing Systems Umashankar Nagarajan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243 Design and Optimization Strategies for Muscle-like Direct Drive Linear Permanent Magnet Motors Bryan P. Ruddy and Ian W. Hunter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 Study of Group Food Retrieval by Ants as a Model for Multi-robot Collective Transport Strategies Spring Berman, Quentin Lindsey, Mahmut Selman Sakar, Vijay Kumar, and Stephen Pratt . . 259 Incremental Sampling-based Algorithms for Optimal Motion Planning Sertac Karaman and Emilio Frazzoli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267 Stochastic Modeling of the Expected Time to Search for an Intermittent Signal Source Under a Limited Sensing Range Dezhen Song, Chang-Young Kim, and Jingang Yi . . . . . . . . . . . . . . . . . . . . . . . 275 Closing the Learning-Planning Loop with Predictive State Representations Byron Boots, Sajid M. Siddiqi, and Geoffrey J. Gordon . . . . . . . . . . . . . . . . . . . . 283 Belief Space Planning Assuming Maximum Likelihood Observations Robert Platt Jr., Russ Tedrake, Leslie Kaelbling, and Tom´as Lozano-P´erez . . . . . . . . . . 291 Motion Planning under Bounded Uncertainty Using Ensemble Control Aaron Becker and Timothy Bretl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299 Remotely Powered Propulsion of Helical Nanobelts Gilgueng Hwang, Sinan Haliyo, and St´ephane R´egnier . . . . . . . . . . . . . . . . . . . . 307 A Molecular Algorithm for Path Self-Assembly in 3 Dimensions Rebecca Schulman and Bernard Yurke . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315

 vii

Preface This volume contains the 40 papers presented at Robotics: Science and Systems (RSS) 2010, held at the University of Zaragoza in Spain, from June 27 to June 30, 2010. A record of 239 paper were submitted to RSS 2010. RSS takes pride in having a rigorous reviewing process: the 218 members of the program committee wrote more than 1, 000 high-quality reviews so that each paper received at least four. The authors were invited to rebut the reviews, and after further discussion between the program committee members and the 16 area chairs, the reviews were finalized. Final acceptance and presentation category (20 poster presentations, 20 podium presentations) was decided during the full-day area chair meeting in Los Angeles, making the acceptance rate of RSS be under 17%. The selected papers cover a wide range of topics in robotics spanning mechanisms, kinematics, dynamics and control, human-robot interaction and human-centered systems, distributed systems, mobile systems and mobility, manipulation, field robotics, medical robotics, biological robotics, robot perception, and estimation and learning in robotic systems. The conference spanned three and a half days. There were four invited talks by leaders in fields that inspire robotics: • Prof. Christopher M. Bishop from Microsoft Research Cambridge and the University of Edinburgh gave a talk titled “Third Generation Machine Intelligence”. • Prof. Deborah M. Gordon from Stanford University gave a talk titled “Interaction Networks as Distributed Algorithms in Ants”. • Prof. Larry Matthies from the Jet Propulsion Laboratory and the University of Southern California gave a talk titled “Robotic Solar System Exploration: Progress and Challenges”. • Prof. Philip H. S. Torr from Oxford Brookes University gave a talk titled “Towards Global Energy Models for Scene Understanding”. There were two early carrer spotlight talks this year, given by rising stars in the robotics community: • Russ Tedrake from Massachusetts Institute of Technology gave a talk titled “Dynamic walking on rough terrain and flying like a bird: a computational approach to exploiting nonlinear dynamics”. • Pieter Abbeel from University of California at Berkeley gave a talk titled “Apprenticeship Learning for High-Performance Robot Control”.

 ix

The Workshop Chair, Kevin Lynch (Northwestern University), together with a group of very competent reviewers, selected twelve workshops that were extremely well attended: more than 260 people attended the eight Sunday workshops, and more than 110 attended the four Monday morning workshops. The workshops were the following: • OMNIVIS Omnidirectional Vision, Camera Networks and Non-classical Cameras organized by Ryad Benosman, Christopher Geyer, and Olivier Koch; • Towards Closing the Loop: Active Learning for Robotics organized by Rub´en Mart´ınezCant´ın, Jan Peters and Andreas Krause; • RGB-D: Advanced Reasoning with Depth Cameras organized by Xiaofeng Ren, Dieter Fox, Jana Kosecka and Kurt Konolige; • Predictive Models in Humanoid Gaze Control and Locomotion organized by Paolo Dario, Alain Berthoz, Jose Santos-Victor and Atsuo Takanishi; • Strategies and Evaluation for Mobile Manipulation in Household Environments organized by Antonio Morales, Mario Prats, Siddhartha Srinivasa and Radu Bogdan Rusu; • Motion Planning: from Theory to Practice organized by Kris Hauser, Ron Alterovitz, Kostas Bekris and Juan Cortes; • Grasp Acquisition: How to Realize Good Grasps organized by Jeff Trinkle, Patrick van der Smagt and Thomas Wimboeck; • Learning for Human-Robot Interaction Modeling organized by Mohamed Chetouani and Adriana Tapus; • Enabling Technologies for Image-Guided Robotic Interventional Procedures organized by Gregory Fischer and Robert Webster; • Representations for Object Grasping and Manipulation in Single and Dual Arm Tasks organized by Danica Kragic, Bruno Siciliano and Vronique Perdereau; • Good Experimental Methodology in Robotics and Replicable Robotics Research organized by Fabio Bonsignorio, John Hallam and Angel del Pobil; • Non-Smooth Contact Modeling in Robotic Simulation and Control organized by Katsu Yamane and Abderrahmane Kheddar.

x

RSS 2010 was a big success thanks to the efforts of many people. We gratefully acknowledge the enormous effort and time spent by the program committee and the 16 area chairs, whose joint expertise covered an extraordinary broad swath of the robotics landscape. The area chairs were: Alin Albu-Schffer (DLR, Germany), Jos´e Castellanos (Universidad de Zaragoza), Peter Corke (CSIRO, Australia), Gregory Dudek (McGill University), Vincent Hayward (Universit´e Pierre et Marie Curie), David Hsu (National University of Singapore), Dan Koditschek (University of Pennsylvania), Jana Kosecka (George Mason University), Danica Kragic (Royal Institute of Technology, Sweden), James Kuffner (Carnegie Mellon University), Jan Peters (Max Planck Institute for Biological Cybernetics), Sanjiv Singh (Carnegie Mellon University), Cyrill Stachniss (University of Freiburg), Russ Tedrake (Massachusetts Institute of Technology), Stefan Williams (University of Sydney) and Katsu Yamane (Disney Research). Our warmest thanks also to the publicity chair, Katherine Kuchenbecker (University of Pennsylvania), who did a great job in attracting the largest yet number of authors and attendees to an RSS conference. The local arrangements chairs, Jos´e Neira, Juan D. Tard´os and Luis Montano (University of Zaragoza), chose an amazing venue and organized everything beautifully, making sure that attendees got the best value for their money. Enormous thanks also to our Webmaster, David Ribas (University of Girona), who designed and maintained an elegant and ergonomic webpage. We also want to thank the conference support staff the Instituto de Investigaci´on en Ingenier´ıa de Arag´on for handling local details, even on a Sunday. Last but not least, we would like to thank the team of twenty three volunteers that were always enthusiastically at hand when something needed to be done. RSS 2010 was possible because of industrial and institutional sponsoring; thanks to Willow Garage, Google, the Spanish Ministry of Science and Innovation, the Regional Government of Arag´on, Heartland Robotics, Aldebaran Robotics, Barrett Technology and Robotnik for providing funds for the general conference. Thanks to Springer for funding the best student paper award, and to Willow Garage for funding the best open source code award. We would also like to thank our technical sponsors: IEEE Robotics and Automation Society, the Association for the Advancement of Artificial Intelligence, the International Foundation of Robotics Research and the Robotics Society of Japan. Finally, we would like to thank again the robotics community for adopting RSS and its philosophy. RSS 2010 had a record attendance of 315 researchers from 25 countries in North America, Europe, Asia and Australia. The attendance rate of 8 persons per presented paper confirms RSS as one of the highest quality single-track mainstream robotics conferences. We enthusiastically look forward to yet more exciting meetings in the years to come. The online version of these proceedings (including color and links) can be found at: http://www.roboticsproceedings.org/rss06/index.html Yoky Matsuoka, University of Washington Hugh Durrant-Whyte, University of Sydney Jos´e Neira, University of Zaragoza July 2010  xi

Organizing Committee General Chair

Yoky Matsuoka, University of Washington

Program Chair

Hugh Durrant-Whyte, University of Sydney

Local Arrangement Co-Chairs

Jos´e Neira, Universidad de Zaragoza Juan Tard´os, Universidad de Zaragoza Luis Montano, Universidad de Zaragoza

Publicity Chair

Katherine Kuchenbecker, University of Pennsylvania

Publications Chair

Jos´e Neira, Universidad de Zaragoza

Workshop Chair

Kevin Lynch, Northwestern University

Web Master

David Ribas, Universitat de Girona

Area Chairs

Alin Albu-Schffer, DLR, Germany Jos´e Castellanos, Universidad de Zaragoza Peter Corke, CSIRO, Australia Gregory Dudek, McGill University Vincent Hayward, Universit´e Pierre et Marie Curie David Hsu, National University of Singapore Dan Koditschek, University of Pennsylvania Jana Kosecka, George Mason University Danica Kragic, Royal Institute of Technology, Sweden James Kuffner, Carnegie Mellon University Jan Peters, Max Planck Institute for Biological Cybernetics Sanjiv Singh, Carnegie Mellon University Cyrill Stachniss, University of Freiburg Russ Tedrake, Massachusetts Institute of Technology Stefan Williams, University of Sydney Katsu Yamane, Disney Research

RSS Foundation Board President

Sebastian Thrun, Stanford University

Directors

Oliver Brock, Technische Universitt Berlin Dieter Fox, University of Washington Lydia Kavraki, Rice University Sven Koenig, University of Southern California John Leonard, Massachusetts Institute of Technology Daniela Rus, Massachusetts Institute of Technology Stefan Schaal, University of Southern California Gaurav Sukhatme, University of Southern California Jeff Trinkle, Rensselaer Polytechnic Institute

Treasurer

Wolfram Burgard, University of Freiburg

Secretary

Nick Roy, Massachusetts Institute of Technology  xiii

Program Committee Abbeel, Pieter Adams, Martin Akella, Srinivas Andrade-Cetto, Juan Antonelli, Gianluca Argyros, Antonis Arras, Kai Oliver Asano, Fumihiko Asfour, Tamim Atkeson, Chris Bagnell, James Bailey, Tim Barkby, Stephen Bennewitz, Maren Bergbreiter, Sarah Bingham, Brian Birchfield, Stan Blaschko, Matthew Borst, Christoph Bosse, Michael Brennan, Sean Brock, Oliver Brooks, Alex Bryson, Mitch Buehler, Martin Burschka, Darius Byl, Katie Carloni, Raffaella Censi, Andrea Chestnutt, Joel Christensen , Henrik Civera, Javier Clark, Jonathan Cowan, Noah Cutkosky, Mark Davison, Andrew Deisenroth, Marc Dellaert, Frank Dissanayake, Gamini Douillard, Bertrand Etienne , Burdet Eustice, Ryan

Everett, Hazel Fairfield, Nathanial Fermuller, Cornelia Ferrie, Frank Fitch, Robert Fox, Dieter Frisoli, Antonio Geyer, Chris Ghrist, Robert Gillespie, Brent Goldman, Daniel Gomez, Manuel Gould, Steve Grizzle, Jessy Grocholsky, Ben Grollman, Dan Grosse-Wentrup, Moritz Gruppen, Rod Guerrero Campo, Jos´e Hafner, Roland Harders, Matthias Hauser, Kris Hirai, Shinichi Hirche, Sandra Hoburg, Warren Hollinger, Geoff Hosoda, Koh Howard, Andrew Huebner, Kai Hyon, sang-ho Iida, Fumiya Inamura, Tetsunari Jadbabaie, Ali Jakuba, Michael Jenkin, Michael Jenkins, Chad Jensfelt, Patric Johnson-Roberson, Matthew Kagami, Satoshi Kajita, Shuuji Kersting, Kristian  xv

Kikuuwe, Ryo Kim, Sangbae Kober, Jens Kolter, Zico Kootstra, Gert Kroemer, Oliver Kuipers, Benjamin Kulic, Dana Kurniawati, Hanna Kyrki, Ville Lacroix, Simon Lampert, Christoph Lane, David LaValle, Steven Lee, Dongheui Leonard, John Leonardis, Alex Lepetit, Vincent Li, Tsai-Yen Lien, Jyh-Ming Likachev, Maxim Lilienthal, Achim Lopes, Manuel L´opez-Nicol´as, Gonzalo MacDonald, Bruce MacLean, Karon Macnab, Chris Mahon, Ian Mansard, Nicolas Mart´ınez-Cant´ın, Rub´en Matsuoka, Yoky Merino, Luis Metta, Giorgio Michael, Nathan Mochiyama, Hiromi Moll, Mark Montesano, Luis Montiel, Jos´e M. M. Morales, Antonio Morimoto, Jun Morse, Bryan Muelling, Katharine

Murillo, Ana Murrieta, Rafael Negahdaripour, Shahriar Neira, Jos´e Nenchev, Dragomir Neumann, Gerhard Newman, Paul Nguyen-Tuong, Duy Niemeyer, Gunter Nishiwaki, Koichi Nuske, Steve O’Kane, Jason Ogata, Tetsuya Okada, Masafumi Olson, Edwin Ott, Christian Oztop, Erhan Palli, Gianluca Papanikolopoulos, Nikolaos Paz, Lina Petillot, Yvan Pfaff, Patrick Piater, Justus Pinies, Pedro Pizarro, Oscar Plagemann, Christian Plaku, Erion Platt, Robert Posner, Ingmar Pradalier, C´edric Pratichizzo, Domenico

Pratt, Jerry Rajan, Kanna Ramos, Fabio Redon, Stephane Rekleitis, Ioannis Riedmiller, Martin Roberts, Jonathan Robuffo Giordano, Paolo Rock, Stephen Rodr´ıguez-Losada, Diego Roman, Chris Roumeliotis, Stergios Roy, Nicholas Sag¨ue´ s, Carlos Sanz, Pedro Saranli, Uluc Saripalli, Srikanth Saxena, Ashutosh Schaal, Stefan Schiele, Bernt Sentis, Luis Shibata, Tomohiro Shimoda, Shingo Shiriaev, Anton Sim, Robert Simeon, Thierry Singh, Surya Singh, Hanumant Smart, Bill Sol`a, Joan Spenko, Matt Spletzer, John

 xvi

Srinivasa, Siddhartha Stasse, Olivier Stilman, Mike Sugihara, Tomomichi Sukhatme, Gaurav Sukkarieh, Salah Tard´os, Juan Taylor, CJ Ting, Jo-Anne Tipaldi, Diego Torres-Mendez, Luz Ude, Ales Ueda, Jun van den Berg, Jur van der Stappen, Frank Vanderborght, Bram Velagapuddi, Pras Venture, Gentiane Villani, Luigi Vona, Marsette Wang, Zhikun Whitcomb, Louis Wingate, David Wyeth, Gordon Yim, Mark Yokokohji, Yasuyoshi Zha, Hong-Bin Zhang, Hong Zillich, Michael Zlot, Robert

Sponsors The organizers of Robotics Science and Systems 2010 gratefully acknowledge the following conference sponsors: • Gold Sponsors:

• Silver Sponsors:

• Bronze Sponsors:

• Awards Sponsors:

• Institutional Sponsors:

 xvii

• Technical Sponsors:

• Organized by:

 xviii

Biophysically,nspired'evelopmentRfa 6Dnd-6wimminJ5Rbot Ryan D. Maladen∗ , Yang Ding† , Paul B. Umbanhowar‡ , Adam Kamor† and Daniel I. Goldman∗†

∗ Bioengineering

Program, † School of Physics, Georgia Institute of Technology, Atlanta, Georgia 30332–0250 of Mechanical Engineering, Northwestern University, Evanston, IL 60208 email: [email protected]

‡ Department

Abstract— Previous study of a sand-swimming lizard, the sandfish, Scincus scincus, revealed that the animal swims within granular media at speeds up to 0.4 body-lengths/cycle using body undulation (approximately a single period sinusoidal traveling wave) without limb use [1]. Inspired by this biological experiment and challenged by the absence of robotic devices with comparable subterranean locomotor abilities, we developed a numerical simulation of a robot swimming in a granular medium (modeled using a multi-particle discrete element method simulation) to guide the design of a physical sand-swimming device built with off-the-shelf servo motors. Both in simulation and experiment the robot swims limblessly subsurface and, like the animal, increases its speed by increasing its oscillation frequency. It was able to achieve speeds of up to 0.3 body-lengths/cycle. The performance of the robot measured in terms of its wave efficiency, the ratio of its forward speed to wave speed, was 0.34±0.02, within 8 % of the simulation prediction. Our work provides a validated simulation tool and a functional initial design for the development of robots that can move within yielding terrestrial substrates.

I. I NTRODUCTION There is a need for robots that can move within complex material like sand, rubble, and loose debris. For example such robots could help locate hazardous chemical leaks [2], function as self propelled inspection devices [3], and search for victims in disaster sites [4, 5, 6]. Limbless robots that use their bodies to move appear better suited to navigate complex terrains than traditional wheeled [7, 8, 9, 10] and legged [11, 12, 13, 14, 15, 16, 17] robots which are often impeded by the size or shape of their appendages which can result in entrapment or failure. Previous terrestrial limbless robots utilized serpentine locomotion to move on the surface of media. Of these, most were tested on rigid surfaces [18, 19, 20, 21, 22] with only a few developed for and tested in unstructured environments [23, 24, 25]. Advances in creating high performing flying and swimming devices [22, 26, 27] in aerial and aquatic domains and wheeled/tracked vehicles on relatively structured terrestrial terrain have occurred mainly because the respective fields of aerodynamics, fluid dynamics, and terramechanics [28, 26] provide accurate models of locomotor-media interaction which are used in turn to design improved wings, fins, wheels, and legs. However, a major hurdle arises when one attempts to design robots to move on and within complex flowing particulate environments (e.g. sand, soil, and leaf-litter) that can display both solid and fluid-like behavior in response to stress. In such materials, comparable and comprehensive val-

1

idated analytic continuum theories at the level of the NavierStokes equations [29] for fluids do not exist. However, it is possible to understand the interaction between the locomotor and the media by using numerical and physical modeling approaches [30, 31, 32]. In the absence of theory, the biological world is a fruitful source of principles of movement that can be incorporated into the design of robots that navigate within complex substrates. Many desert organisms like scorpions, snakes, and lizards burrow and swim effectively in sand [33, 34, 35, 36, 37] to escape heat and predators, and hunt for prey [38, 39]. It has been hypothesized that many of these animals have evolved morphological adaptations like marked body elongation and limb reduction to deal with deformable terrain [40, 41]. Our recent high speed x-ray imaging study investigating the subsurface locomotion of the sandfish Scincus scincus, a small (∼ 10FPORQJVQRXWWDLOWLS OizardWKDWLnhabitsWhe Saharan desert [1] (Fig. 1), reveals that once within the media the animal no longer uses limbs for propulsion but “swims” forward by propagating a sinusoidal traveling wave posteriorly from head to tail. Motivated by the subsurface locomotion of the sandfish, the present work utilizes a numerical simulation of a sandfish inspired undulator as a design tool to build an appendageless sand-swimming device. The robot is driven by a simple open loop controller which, like the animal kinematics, varies the joint position trajectories to create a sinusoidal wave that travels posteriorly along the device. The robot swims within a model laboratory granular medium of plastic particles and displays locomotion features similar to the organism and predicted by the numerical robot simulation. II. P REVIOUS WORK A. Biological(xperiment The biological experiments presented in [1] model the subsurface undulatory motion of the sandfish with a posteriorly traveling single-period sinusoidal wave 2π (1) (x + vw t) λ with x the position along the sandfish, y the body displacement from the midline of the animal, A the amplitude, λ the wavelength and vw = f λ the wave speed where f is the wave frequency. The spatial characteristics, A and λ, did y = A sin

A

B

C

Fig. 1. (A) The sandfish Scincus scincus, a sand-swimming lizard that inhabits the Saharan desert, (B) burying into granular media (0.3 mm spherical glass beads), and (C) swimming subsurface where the x-ray image shows the body (light area) and opaque markers fixed to limbs and midline. Red dashed line indicates tracked midline.

not vary significantly with media preparation and their ratio was approximately 0.2 implying that the animal increased its forward velocity by increasing its oscillation frequency. A measure of undulatory performance is the wave efficiency, η, the ratio between the forward speed of the swimmer, vx , and the velocity of the wave traveling down its body, vw , or equivalently the slope of the velocity-frequency relationship for velocity measured in wavelengths per second. Typical wave efficiencies of undulatory organisms moving in fluids at low Reynolds number (such as nematodes in water) are 0.25 [42, 43, 44], whereas η ≈ 0.8 − 0.9 for organisms undulating (creeping) along solid – air interfaces [45, 46, 47]. Locomotion with η = 1 is equivalent to movement within a rigid tube. For the sandfish swimming in glass beads, η ≈ 0.5 independent of particle size and media preparation (i.e. packing density). B. Resistive Force Theory for*ranulaU0edia An empirical resistive force theory (RFT) was developed to predict wave efficiency η for undulatory subsurface granular locomotion [1]. The RFT, inspired by theory used to predict swimming speeds of microorganisms in fluids [42], partitions the body of the organism into infinitesimal segments each of which generates thrust and experiences drag when moving through the medium. These segmental forces are integrated over the entire body, and, by setting the net forward force to zero (assuming a constant average velocity), η is solved for numerically.

2

Unlike fluids, in granular media no validated theory exists in the regime relevant to sand-swimming to estimate the force on individual segments moving through the medium. Previously, Maladen et al. [1] obtained these forces empirically by dragging a rod (representative segment) through the media the animal was tested in. With these forces as input and by propagating a sinusoidal traveling wave along the body, the RFT shows that translational motion within granular media without limb use is possible. Also, the RFT accurately predicts that the sandfish swims with η ≈ 0.5 within a granular media of 0.3 mm glass particles (comparable in size and density to desert sand [48]). While the RFT qualitatively describes some features of sand-swimming, it is based on several assumptions: e.g. the measured drag force on a rod is representative of the average force on a segment of the sandfish, the forces generated by a segment are localized, and the center of mass of the animal does not oscillate laterally. Since the assumptions of the RFT have not been rigourously tested and applying the RFT to different treatments (particle friction, particle size, body design, etc.) require force laws to be measured for each condition, we instead use numerical simulation techniques as a general robotic design tool. A numerical simulation approach, once validated against experiment, can provide an understanding of body generated drag and thrust forces from the particle perspective and can be used to generate empirical drag laws for input into the RFT. Our numerical simulation is a flexible design tool that accurately predicts robot performance and allows easy variation of physical and design parameters such as particle-particle friction and number of segments. III. N UMERICAL S IMULATION OF S AND -S WIMMING ROBOT A. Development and9alidation To design a sand-swimming robot, we developed a numerical simulation of a laboratory scale device with a finite number of discrete, rigid segments to test if it could swim within granular medium. The simulation couples a numerical model of the robot to a model of the granular medium. We model the granular material using a multi-particle discrete element method (DEM) simulation [30]. To compute the robot-particle and particle-particle interaction forces we calculate the normal force [49], Fn , and the Coulombic tangential force, Fs , acting at each contact with Fn = kδ 3/2 − Gn vn δ 1/2 Fs = μFn ,

(2)

where δ is the virtual overlap between particles or between particle and robot segment, vn is the normal component of relative velocity, and k = 2 × 105 kg s−2 m−1/2 and Gn = 5 kg s−1 m−1/2 are the hardness and viscoelastic constant. μ quantifies the particle-particle (μpp = 0.08) or body-particle (μbp = 0.27) friction coefficient depending on which elements are in contact. μbp was measured between the robot skin

Acceleration (g)

and plastic particles used in the physical experiments. To in a horizontal plane at fixed depth in this study, and there reduce the required torque in the physical experiments and to was excellent agreement between experiment and simulation decrease the computational time, we used a granular medium in preliminary studies. The simulated robot was sized for easy composed of 4.7×105 spherical plastic particles with diameter testing of the corresponding physical device in the laboratory. 6 mm and density 1.03 ± 0.04 g/cm3 in experiment, and Since the sandfish does not use its limbs to move subsurface 3.2 × 105ELGLVSHUVH VSKHUHPL[WXUHPP 6and RFT had shown that body undulation was sufficient for and density 1.06 g/cm3 in simulation. The 35 particle deep propulsion [1], the simulated robot did not include limbs. bed of particles in experiment and 24 particle deep bed of No tapering along the device was considered. The simulated particles in simulation were held in a container with horizontal robot consisted of 49 cuboidal segments interconnected and cross section of 188 × 62 particle diameters. To validate the actuated by virtual motors (vertical cylinders) of the same simulated medium and obtain the values of μpp , k, and Gn height (Fig. 3). Depending on the number of segments (N ) given above, we dropped an aluminum ball (diameter 6.35 cm employed, every 48/N motor was driven with an open loop and mass 385 g) into the plastic particles with varying impact signal to generate a sinusoidal wave traveling posteriorly from velocity (0.5 − 3 m/s) in both experiment and simulation and head to tail while the remaining motors were immobilized to set grain interaction parameters to best match the measured form a straight segment of length 48 / N cm. To approxiand simulated penetration force during the impact collision as mate a sinusoidal traveling wave, the angle between adjacent a function of time (Fig. 2). With parameters determined from segments is modulated using impact at v = 1.4 m/s, the force profile fit well at other impact velocities. In additional experiments, we directly measured μpp β(i, t) = β0 ξ sin(2πξi/N − 2πf t), (3) and the coefficient of restitution (determined by Gn with fixed th k) for the plastic particles and found them to be within 5% and with β(i, t) the motor angle of the i motor at time t, β0 10% of the fitted values respectively. For simplicity we used the angular amplitude which determines A/λ, ξ the number the same normal force parameters for both particle-particle of wavelengths along the body (period), and N the number of motors. and body-particle interactions. WM integrates the equations of motion of the coupled links and the DEM calculates the resultant forces from both the particle-particle and body-particle interactions. For each time 2 cm step, the net force from particles on each segment is passed to 4 WM, and velocity and position information transferred back to DEM. Roll and pitch were not modeled.

A

2

motor i-1

D

βi

3 cm

motor i+1

motor i

B

0.05 Time (s)

Experiment Simulation 0.1

[

0

[

0

Fig. 2. Validation of the multi-particle discrete element method (DEM) simulation of the granular medium using measured acceleration of a sphere during vertical impact after free-fall. Acceleration vs. time in simulation (blue dashed trace) and experiment (red solid trace) agree well. The impact velocity for this representative run is 1.4 m/s. Acceleration is in units of g, the acceleration due to gravity. (Left inset) Aluminium ball instrumented with accelerometer resting on 6 mm plastic particles. (Right inset) Ball and particles in simulation.

To model the sand-swimming device we used the commercial multi-body simulator software package Working Model (WM) 2D (Design Simulation Technologies). Modeling the device in a 2D simulation environment is sufficient to capture the dynamics since the sand-swimming robot moves roughly

3

C

[

Fig. 3. Simulation of a sand-swimming robot. (A,B) Side and top view of the robot modeled with 49 inter-connected motor segments and one head segment. The angle between adjacent motors (βi ) is modulated using Eqn. 3 to reproduce the sandfish’s sinusoidal traveling wave kinematics. (C) Top view of the device submerged in 6 mm particles with particles above the robot rendered transparent. (D) Rendering of the simulated robot for the same parameters used in robot experiment (see Fig. 4). The brackets ( [ ) indicate a single robot segment.

Using Eqn. 3 the simulated robot with 7 total segments moved forward within 6 mm plastic particles and increased its forward speed linearly with oscillation frequency (Fig. 7). The wave efficiency was η = 0.36 ± 0.02, less than that of the sandfish lizard. Motivated by these results we built a physical instantiation of the scaled model. IV. S AND S WIMMING ROBOT

5cm

B tail A

head

A. Design and Control The basic mechanical design of our device was adapted from previously developed snake robots [25] which consisted of repeated modules (motors) each with a single joint that permit angular excursions in a plane and connected via identical links. In our design, each module consists of a servomotor attached to an aluminium bracket and is connected to adjacent motors via the brackets. The wire bundle that routes power and control signals to each motor was run atop each module over the length of the device and strain relieved at the last (tail) segment. For convenience and to maintain a reasonable size, our device employed 6 standard size servomotors and a dummy segment (the head) with the same weight and form factor as the motor segments for a total of 7 segments (Fig. 4A). The simulation found that the peak torque required to swim subsurface at a depth of 4 cm was 0.7 N m. To verify this finding we dragged an object with the same form factor as a motor through the 6 mm plastic medium at 0.25 m/s. The measured force at a depth of 4 cm was 3.2 N. Since the maximum torque occurs at the middle motor (0.23 m to either end) we estimated the maximum possible total force along an effective segment extending from the middle servo to either the tail or the head (length 0.23 m) to be 18 N with a corresponding maximum torque of 2.0 Nm. We selected a servomotor that exceeded both torque estimates, see Table I. Servomotors are powered in parallel from a 7.4 V, 30 A supply. The pulse width based control signal for each motor is generated in LabVIEW using Eqn. 3 as a multiplexed signal, output from a PCI-card (NI-6230), and connected to the clock input of a decade counter (CD4017BC) which functions as a demultiplexer and distributes a control pulse to each motor every 20 ms. Since the robot operates in a granular medium it is critical to encase it in a material that prevents particles from getting between the motor segments but allows the device to easily undulate. After testing several materials we found that a 2layer encasement consisting of an outer Lycra spandex sleeve with a single seam (located at the top of the device) enclosing an inner thin latex sleeve that fit tautly around the motors was satisfactory (Fig. 4). B. Experimental Methods We tested the robot in a container of the same dimensions as used in the robot simulations and filled with 6 mm plastic particles prepared in an as-poured state [50]. Overhead video (100 fps) was collected for each condition tested. To facilitate subsurface tracking the first and last module were fixed with

4

head

tail

B

tail

head

C Fig. 4. Prototype of the sand-swimming robot. (A) basic construction (servomotors and aluminium brackets with power wires running along the top of the device). The robot has a double layer skin: (B) tight fitting thin latex inner layer, and (C) Lycra spandex outer layer. Balls on narrow masts on the head and tail segments allow subsurface motion tracking.

a mast with a visible marker. The wire bundle was run up the mast on the last segment and tethered above the container. The kinematics of the subsurface motion of the robot were also obtained using x-ray imaging for a representative condition (f = 0.25 Hz, A/λ = 0.2), see Fig. 5A-C. For each test the top of the robot was submerged 4 cm below the surface and the surface leveled. Due to the servomotor angular velocity limits the maximum oscillation frequency was 1 Hz. For each frequency, 1 − 2 cycles of motion were collected.

0.48 × 0.028 × 0.054 m3 0.83 kg HSR-5980SG 2.94 N m 6 7

6 mm plastic beads segment S1

TABLE I P HYSICAL ROBOT CHARACTERISTICS

A

B

B

C

35

tail

S4

S1 S1

A 65 y (cm)

Dimensions Mass Motor Motor Torque Number of Motors Total Segments

motion

S7

motion head

D

3 cm

motion

E

F S4

S1

tail

S7

segment S7

head

2 cm

Fig. 5. Subsurface swimming in experiment and simulation. (A-C) Sequential x-ray images of the robot swimming in 6 mm particles, and (D-F) robot swimming in simulation. Segments from head to tail are denoted as S1 to S7.

C. Robot Performance To calibrate the device we placed it on a rigid surface and used video to track the position of the segments from which we determined the mapping between the maximum relative segment angle β0 (Eqn. 3) and A/λ. Within the granular material, the forward velocity of the device monotonically increased with increasing oscillation frequency (Fig. 7) for A/λ = 0.2 and a single period wave. The slope of this relationship (η) was 0.34 ± 0.02. For the same parameters the simulation predicted η = 0.36 ± 0.02. V. D ISCUSSION Like the sandfish, the robot swims within granular media by propagating a traveling sinusoidal wave posteriorly from head to tail without limb use. The physical device demonstrates that subsurface locomotion in granular media using a relatively low degree of freedom device and a open loop control scheme is possible. However, the robot does not move forward as fast (normalized by body-length) or with the same wave efficiency as the animal. In the biological experiments, η for a range of granular material preparations and bead size was approximately 0.5. The robot in both experiment and simulation performed below this value. We hypothesized that the number of segments (for a fixed length device) affected both η and the forward speed of the device. Increasing the number of segments in the robot simulation caused the device to move forward faster and with greater wave efficiency until N ∼ 15 where η

5

5 cm

0 0 x (cm) 15

Fig. 6. Subsurface swimming in experiment and simulation. (A) Robot submerged in a container filled with 6 mm plastic particles. Masts with spherical markers are attached to the first and last module. (B) Kinematics of the first and last segment of the robot in experiment (green circles) and simulation (blue triangles).

plateaued (Fig. 8). Interestingly, the maximum η ≈ 0.5 is the same as measured in the animal experiment. We utilized our previously developed RFT to predict the performance of the sand-swimming device with parameters set to match those for the plastic particles used in the robot experiment. We estimated η = 0.56 for a smooth profiled undulator which corresponds to the numerical robot simulation prediction for N > 15 (gray band, Fig. 8). Increasing N allowed the device to better match a sinusoidal wave and increased η This suggests that deviation from the smooth form of a traveling sinusoidal wave reduces performance. A seven segment robot operates below the minimum N required to achieve maximum η. As a design criterion, N is important when the length of the device is fixed as increasing the number of motors beyond the critical N requires motors with smaller dimensions but capable of producing the same torque. We used the numeric robot simulation to measure the time varying torque required to move within the medium. As expected, the torque was approximately sinusoidal for all motors and the torque amplitude generated by the central motors (3 and 4) was larger than the torque from the motors nearest the ends, see Fig. 9. As noted earlier, the maximum torque in the simulation of 0.7 N m was well below the maximum of the motors used in experiment (see Table 1).

0.7 Torque (N m)

η = 0.36 0.3 η = 0.34

0

0

Frequency (Hz)

A

0

−0.7 0 0.7

Torque amplitude (Nm)

-1 Speed (vx /λ) (s )

Simulation Experiment

1

Fig. 7. Forward velocity vs. oscillation frequency for the robot in experiment (green circles) and simulation (blue triangles) (A/λ = 0.2). The slope of the dashed (simulation) and solid (experiment) fit lines gives the wave efficiency η.

2

4

6

B

0.35

tail 1

0

Time (s)

head 3

6

5

6

4

3 Motor

2

1

Fig. 9. Motor torque for the simulated 7 segment, 6 motor robot (f = 1 Hz) (A) varying with time. (B) Torque amplitude vs. motor position; orange (solid curve), green (dotted curve) , and black (dash-dot curve) correspond to motor 6 (tail), 4, and 1 (head) with motor position 1 denoting segment number 2 in Fig. 5 and 6.

0.7

η

0.35

N= 5

0 0

10

N= 15

N= 48

20 30 40 Number of Segments

50

Fig. 8. Wave efficiency increases with number of segments for a fixed length robot in simulation (blue dashed curve)(f = 1 Hz and A/λ = 0.2). The red, black, and cyan triangles correspond to 5, 15, and 48 segment robots respectively. The green square corresponds to the seven segment physical robot, and the grey line indicates η predicted by the RFT solved for a continuous body profile (see text for details).

Also, the fluctuations in torque at frequencies higher than the oscillation frequency of the robot were small in comparison to the torque amplitude. VI. F UTURE W ORK A sand swimming robot combined with a proven simulation tool opens many avenues for further research. Of immediate interest is testing the RFT prediction that an optimal spatial form (ratio of amplitude to wavelength) maximizes forward speed of an undulatory sand swimmer [1]. The effect of the predicted optimal kinematics can also be evaluated by measuring the mechanical cost of transport. In conjunction with the numerical simulation the robot can test the effect of motion profiles (wave shapes) on performance. Since the

6

sandfish uses the same kinematics to move in a variety of media, duplicating the animals control methods and sensing modalities in a robot could lead to more effective locomotion. The sandfish has a non-trivial shape which suggests changing the morphological characteristics of the robotic device. For example, the cross sectional shape of the sandfish (flat belly and rounded top) have been hypothesized to aid rapid burial into granular media [51]. Our simulated and physical robot can be used to explore the influence of this morphology along with body taper on performance. The robotic simulation can also tune parameters like skin friction and body compliance to identify optimal values which could then be tested with our robot. VII. C ONCLUSION Motivated by biological experiments revealing rapid subsurface sand-swimming in the sandfish lizard, we have used numerical simulation as a design tool to build an undulatory sand-swimming device. We used our robot simulation to test whether a device with a finite number of segments (7) could advance using a simple open loop (traveling wave sinusoid) control scheme and calculated the motor torque requirements for the robot. We then built and tested a prototype of the device to validate the biological observations and predictions from the RFT [1] and simulations that limbless body undulations were sufficient to propel the robot forward. Our findings show that the device can swim, and that it translates faster by increasing its oscillation frequency just as the sandfish does. The design tools (numerical model and robot) we developed can generate testable hypotheses of neuromechanical control [52] and improve our understanding of how organisms exploit the solid and fluid-like properties of granular media, enabling the

construction of robots that can locomote effectively within complex environments. ACKNOWLEDGMENTS We thank Nick Gravish and Chen Li for help with the physics experiments, and we thank Daniel Cohen and Andrew Slatton for help with development of the numerical simulation. We also acknowledge our funding from The Burroughs Wellcome Fund Career Award at the Scientific Interface, NSF Physics of Living Systems grant PHY-0749991, and the Army Research Laboratory (ARL) Micro Autonomous Systems and Technology (MAST) Collaborative Technology Alliance (CTA) under cooperative agreement number W911NF08 − 2 − 0004. R EFERENCES [1] R. Maladen, Y. Ding, C. Li, and D. Goldman, “Undulatory Swimming in Sand: Subsurface Locomotion of the Sandfish Lizard,” Science, vol. 325, no. 5938, p. 314, 2009. [2] C. Humphrey and J. Adams, “Robotic Tasks for Chemical, Biological, Radiological, Nuclear and Explosive Incident Response,” Advanced Robotics, vol. 23, no. 9, pp. 1217–1232, 2009. [3] J. McKean, S. Buechel, and L. Gaydos, “Remote sensing and landslide hazard assessment,” Photogrammetric engineering and remote sensing, vol. 57, no. 9, pp. 1185–1193, 1991. [4] G. Metternicht, L. Hurni, and R. Gogu, “Remote sensing of landslides: An analysis of the potential contribution to geo-spatial systems for hazard assessment in mountainous environments,” Remote sensing of Environment, vol. 98, no. 2-3, pp. 284–303, 2005. [5] A. Ashcheulov, I. Gutsul, and V. Maevski, “Device for monitoring the radiation temperature in coal mines,” Journal of Optical Technology, vol. 67, no. 3, p. 281, 2000. [6] W. Marcus, C. Legleiter, R. Aspinall, J. Boardman, and R. Crabtree, “High spatial resolution hyperspectral mapping of in-stream habitats, depths, and woody debris in mountain streams,” Geomorphology, vol. 55, no. 1-4, pp. 363–380, 2003. [7] R. Ritzmann, R. Quinn, and M. Fischer, “Convergent evolution and locomotion through complex terrain by insects, vertebrates and robots,” Arthropod structure and development, vol. 33, no. 3, pp. 361–379, 2004. [8] R. Siegwart, P. Lamon, T. Estier, M. Lauria, and R. Piguet, “Innovative design for wheeled locomotion in rough terrain,” Robotics and Autonomous systems, vol. 40, no. 2-3, pp. 151–162, 2002. [9] J. Wong, “On the study of wheel-soil interaction,” Journal of Terramechanics, vol. 21, no. 2, pp. 117–131, 1984. [10] P. Arena, P. Di Giamberardino, L. Fortuna, F. La Gala, S. Monaco, G. Muscato, A. Rizzo, and R. Ronchini, “Toward a mobile autonomous robotic system for Mars exploration,” Planetary and Space Science, vol. 52, no. 1-3, pp. 23–30, 2004. [11] C. Li, P. B. Umbanhowar, H. Komsuoglu, D. E. Koditschek, and D. I. Goldman, “Sensitive dependence of the motion of a legged robot on granular media,” Proceedings of the National Academy of Science, vol. 106, no. 9, pp. 3029–3034, 2009. [12] C. Li, P. Umbanhowar, H. Komsuoglu, and D. Goldman, “The effect of limb kinematics on the speed of a legged robot on granular media,” Experimental Mechanics, pp. 1–11, 2010. [13] U. Saranli, M. Buehler, and D. Koditschek, “RHex: A simple and highly mobile hexapod robot,” The International Journal of Robotics Research, vol. 20, no. 7, p. 616, 2001. [14] R. Playter, M. Buehler, and M. Raibert, “BigDog,” in Unmanned Ground Vehicle Technology VIII, ser. Proceedings of SPIE, D. W. G. Grant R. Gerhart, Charles M. Shoemaker, Ed., vol. 6230, 2006, pp. 62 302O1– 62 302O6. [15] A. Hoover, E. Steltz, and R. Fearing, “RoACH: An autonomous 2.4 g crawling hexapod robot,” in IEEE Intelligent Robots and Systems Conference Proceeding, 2008, pp. 26–33. [16] A. Saunders, D. I. Goldman, R. J. Full, and M. Buehler, “The rise climbing robot: body and leg design,” in Unmanned Systems Technology VIII, G. R. Gerhart, C. M. Shoemaker, and D. W. Gage, Eds., vol. 6230. SPIE, 2006, p. 623017.

7

[17] S. Kim, J. E. Clark, and M. R. Cutkosky, “iSprawl: Design and tuning for high-speed autonomous open-loop running,” International Journal of Robotics Research, vol. 25, no. 9, 2006. [18] S. Hirose and A. Morishima, “Design and control of a mobile robot with an articulated body,” The International Journal of Robotics Research, vol. 9, no. 2, p. 99, 1990. [19] Y. Shan and Y. Koren, “Obstacle accommodation motion planning,” in Intelligent autonomous systems, IAS–3: proceedings of the international conference, Pittsburgh, Pennsylvania, February 15-18, 1993. Ios Pr Inc, 1993, p. 94. [20] R. Desai, C. Rosenberg, J. Jones, and I. Inc, “Kaa: an autonomous serpentine robot utilizes behavior control,” in Proceedings of the 1995 International Conference on Intelligent Robots and Systems, Pittsburgh. Citeseer, 1995. [21] B. Klaassen and K. Paap, “GMD-SNAKE 2: a snake-like robot driven by wheels and a method for motion control,” in Proceedings from the IEEE International Conference On Robotics And Automation, vol. 4, 1999, pp. 3014–3019. [22] A. Crespi, A. Badertscher, A. Guignard, and A. Ijspeert, “Swimming and crawling with an amphibious snake robot,” in IEEE International Conference On Robotics And Automation, vol. 3. Citeseer, 2005, p. 3024. [23] G. Chirikjian and J. Burdick, “The kinematics of hyper-redundant robot locomotion,” IEEE transactions on Robotics and Automation Proceedings, vol. 11, no. 6, pp. 781–793, 1995. [24] H. Choset, J. Luntz, E. Shammas, T. Rached, D. Hull, and C. Dent, “Design and motion planning for serpentine robots,” in Proceedings of SPIE, vol. 3990, 2000, p. 148. [25] K. Dowling, “Limbless locomotion: learning to crawl,” in IEEE International Conference on Robotics and Automation Proceedings, vol. 4, 1999. [26] H. Kim, D. Shim, and S. Sastry, “Flying robots: modeling, control and decision making,” in IEEE International Conference on Robotics and Automation Proceedings., vol. 1, 2002. [27] J. Zufferey, A. Klaptocz, A. Beyeler, J. Nicoud, and D. Floreano, “A 10-gram vision-based flying robot,” Advanced Robotics, vol. 21, no. 14, pp. 1671–1684, 2007. [28] M. Bekker, Theory of Land Locomotion. The University of Michigan Press, 1956. [29] D. Tritton, Physical Fluid Dynamics. Oxford University Press, 1989. [30] D. C. Rapaport, The Art of Molecular Dynamics Simulation, 2nd ed. Cambridge University Press, 2004. [31] K. Gerritsen, A. van den Bogert, and B. Nigg, “Direct dynamics simulation of the impact phase in heel-toe running,” Journal of Biomechanics, vol. 28, no. 6, pp. 661–668, 1995. [32] A. Mouazen and M. Nemenyi, “Tillage tool design by the finite element method: Part 1. finite element modelling of soil plastic behaviour,” Journal of Agricultural Engineering Research, vol. 72, no. 1, pp. 37–51, 1999. [33] F. Meysman, J. Middelburg, and C. Heip, “Bioturbation: a fresh look at Darwin’s last idea,” Trends in Ecology & Evolution, vol. 21, no. 12, pp. 688–695, 2006. [34] R. Seymour, P. Withers, and W. Weathers, “Energetics of burrowing, running, and free-living in the Namib Desert golden mole (Eremitalpa namibensis),” Journal of Zoology, vol. 244, no. 01, pp. 107–117, 2001. [35] K. Dorgan, S. Arwade, and P. Jumars, “Burrowing in marine muds by crack propagation: kinematics and forces,” Journal of Experimental Biology, vol. 210, no. 23, p. 4198, 2007. [36] L. Fielden, “Home range and movement of the Namib Desert golden mole, Eremitalpa granti namibensis (Chrysochloridae).” Journal of Zoology, vol. 223, no. 4, pp. 675–686, 1991. [37] S. Hawke and R. Farley, “Ecology and behavior of the desert burrowing cockroach, Arenivaga sp. (Dictyoptera, Polyphagidae),” Oecologia, vol. 11, no. 3, pp. 263–279, 1973. [38] E. Arnold, “Identifying the effects of history on adaptation: origins of different sand-diving techniques in lizards,” J. Zool., Lond, vol. 235, pp. 351–388, 1995. [39] C. White, “The influence of foraging mode and arid adaptation on the basal metabolic rates of burrowing mammals,” Physiological and Biochemical Zoology, vol. 76, no. 1, pp. 122–134, 2003. [40] C. Gans, “Tetrapod limblessness: evolution and functional corollaries,” Integrative and Comparative Biology, vol. 15, no. 2, p. 455, 1975. [41] S. Carranza, E. Arnold, P. Geniez, J. Roca, and J. Mateo, “Radiation, multiple dispersal and parallelism in the skinks, Chalcides and Sphenops

[42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52]

(Squamata: Scincidae), with comments on Scincus and Scincopus and the age of the Sahara Desert,” Molecular phylogenetics and evolution, vol. 46, no. 3, pp. 1071–1094, 2008. J. Gray and G. Hancock, “The propulsion of sea-urchin spermatozoa,” Journal of Experimental Biology, vol. 32, no. 4, p. 802, 1955. N. Cohen and J. Boyle, “Swimming at low Reynolds number: a beginners guide to undulatory locomotion,” Contemporary Physics, vol. 51, no. 2, pp. 103–123, 2010. S. Berri, J. Boyle, M. Tassieri, I. Hope, and N. Cohen, “C. elegans locomotion: a unified multidisciplinary perspective,” BMC Neuroscience, vol. 10, no. Suppl 1, p. P16, 2009. J. Gray and H. Lissmann, “The locomotion of nematodes,” Journal of Experimental Biology, vol. 41, no. 1, p. 135, 1964. B. Jayne, “Kinematics of terrestrial snake locomotion,” Copeia, pp. 915– 927, 1986. J. Gray, “The mechanism of locomotion in snakes,” Journal of Experimental Biology, vol. 23, no. 2, p. 101, 1946. R. A. Bagnold, The Physics of Blown Sand and Desert Dunes. Methuen and Co. Ltd., 1954. J. Lee and H. Herrmann, “Angle of repose and angle of marginal stability: molecular dynamics of granular particles,” Journal of Physics A: Mathematical and General, vol. 26, pp. 373–383, 1993. R. Nedderman, Statics and kinematics of granular materials. Cambridge Univ. Pr., 1992. W. Mosauer, “Adaptive convergence in the sand reptiles of the Sahara and of California: a study in structure and behavior,” Copeia, pp. 72–78, 1932. K. Nishikawa, A. A. Biewener, P. Aerts, A. N. Ahn, H. J. Chiel, M. A. Daley, T. L. Daniel, R. J. Full, M. E. Hale, T. L. Hedrick, A. K. Lappin, T. R. Nichols, R. D. Quinn, R. A. Satterlie, and B. Szymik, “Neuromechanics: an integrative approach for understanding motor control,” Integrative and Comparative Biology, vol. 47, no. 1, pp. 16–54, 2007.

8

3DVVLYH 7RUTXH 5HJXODWLRQ LQ DQ 8QGHUDFWXDWHG )ODSSLQJ :LQJ 5RERWLF ,QVHFW 3 6 6UHHWKDUDQ DQG 5 - :RRG 6FKRRO RI (QJLQHHULQJ DQG $SSOLHG 6FLHQFH +DUYDUG 8QLYHUVLW\ &DPEULGJH 0DVVDFKXVHWWV  SUDWKHHY#SRVWKDUYDUGHGX

DHURPHFKDQLFDO SODWIRUP WR DFKLHYH D OLIW WR ZHLJKW UDWLR JUHDWHU WKDQ XQLW\ +RZHYHU WKH EHQHÀWV RI XQGHUDFWXDWLRQ DQG SDVVLYH FRP SOLDQFH FDQ H[WHQG EH\RQG VLPSOH UHGXFWLRQ RI PHFKDQLFDO FRPSOH[LW\ LQ SDUWLFXODU IRU GHYLFHV LQ ZKLFK WKH GLVWULEXWLRQ RI IRUFHV DQG WRUTXHV LV RI IXQGDPHQWDO LPSRUWDQFH $ XELTXL WRXV H[DPSOH LV WKH DXWRPRELOH GLIIHUHQWLDO DQ XQGHUDFWXDWHG PHFKDQLVP FRPPRQO\ XVHG WR GLVWULEXWH HQJLQH SRZHU WR WZR ZKHHOV 7KH GLIIHUHQWLDO LQFRUSRUDWHV DQ DGGLWLRQDO GHJUHH RI IUHHGRP q2 WR EDODQFH WKH WRUTXH GHOLYHUHG WR HDFK ZKHHO VHH )LJ   7KH GLIIHUHQWLDO IXQGDPHQWDOO\ RSHUDWHV RQ ZKHHO WRUTXHV LQVWHDG RI URWDWLRQV DLGHG E\ SDVVLYH PHFKDQLVPV WKH ZKHHOV FDQ URWDWH DORQJ FRPSOH[ UHODWLYH WUDMHFWRULHV PDLQWDLQLQJ WUDFWLRQ RQ WKH JURXQG ZLWKRXW FORVHG ORRS DFWLYH FRQWURO

$EVWUDFW³5HFHQW GHYHORSPHQWV LQ PLOOLPHWHUVFDOH IDEULFDWLRQ SURFHVVHV KDYH OHG WR UDSLG SURJUHVV WRZDUGV FUHDWLQJ DLUERUQH ÁDSSLQJ ZLQJ URERWV EDVHG RQ 'LSWHUDQ WZRZLQJHG LQVHFWV 3UHYLRXV ZRUN WR UHJXODWH UHDFWLRQ IRUFHV DQG WRUTXHV JHQHUDWHG E\ WZR ÁDSSLQJ ZLQJV KDV ODUJHO\ IRFXVHG RQ ZLQJ WUDMHFWRU\ FRQWURO $Q DOWHUQDWLYH DSSURDFK LQWURGXFHV DGGLWLRQDO GHJUHHV RI IUHHGRP WR WKH ZLQJ ÁDSSLQJ PHFKDQLVP WR SDVVLYHO\ UHJXODWH WKHVH IRUFHV DQG WRUTXHV 7KH UHVXOWLQJ ¶PHFKDQLFDOO\ LQWHOOLJHQW· GHYLFHV FDQ H[HFXWH ZLQJ WUDMHFWRU\ FRUUHFWLRQV WR UHDOL]H GHVLUHG ERG\ IRUFHV DQG WRUTXHV ZLWKRXW WKH LQWHUYHQWLRQ RI DQ DFWLYH FRQWUROOHU 7KLV SDSHU GHVFULEHV DQ LQVHFWVFDOH Á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

q1

, , 1752'8&7,21 $GYDQFHV LQ PLOOLPHWHU VFDOH IDEULFDWLRQ SURFHVVHV KDYH HQDEOHG UDSLG SURJUHVV WRZDUGV WKH GHYHORSPHQW RI ÁDSSLQJ ZLQJ PLFUR DLU YHKLFOHV ):0$9V ZLWK V\VWHP PDVV RQ WKH RUGHU RI PJ >@ +RZHYHU ÁLJKW VWDELOLW\ DQG FRQWURO PHFKDQLVPV IRU WKHVH PDVV DQG SRZHU OLPLWHG V\VWHPV UHPDLQ DFWLYH DUHDV RI UHVHDUFK ,QYHVWLJDWLRQ LQWR WKH DHURG\QDPLFV RI ELRORJLFDO LQVHFW ÁLJKW KDV SURGXFHG DSSUR[LPDWH DHURG\QDPLF PRGHOV DOORZLQJ FRPSXWDWLRQDOO\ LQH[SHQVLYH SUHGLFWLRQ RI DHURG\QDPLF IRUFHV DQG WRUTXHV IURP ZLQJ WUDMHFWRULHV >@ >@ $FFRUGLQJO\ UHVHDUFK LQWR WUDQVPLVVLRQ DQG FRQWURO PHFKDQLVPV RI Á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

q2 Aq1 + Bq2

Aq1 – Bq2

to right

wheel

heel

to left w

)LJ  $ FDU GLIIHUHQWLDO EDODQFHV RXWSXW WRUTXHV XVLQJ DQ XQGHUDFWXDWHG PHFKDQLVP 'HJUHH RI IUHHGRP q1 UHFHLYHV HQJLQH WRUTXH ZKLOH q2 LV XQDFWXDWHG 3UHYLRXV ZRUN KDV LQWURGXFHG WKH FRQFHSW RI 3DVVLYH $HURPHFKDQLFDO 5HJXODWLRQ RI ,PEDODQFHG 7RUTXHV 3$5,7\ LQ WKH FRQWH[W RI LQVHFWVFDOH ):0$9 GHVLJQ >@ (PERG\LQJ WKH 3$5,7\ FRQFHSW WKH ¶'UDJ 3$5,7\· LV DQ XQGHUDFWXDWHG WZR GHJUHH RI IUHHGRP ):0$9 WUDQVPLVVLRQ WKDW DQDORJRXV WR DQ DXWRPRELOH GLIIHUHQWLDO SDVVLYHO\ GLVWULEXWHV SRZHU IURP D VLQJOH DFWXDWRU WR EDODQFH WRUTXHV GHOLYHUHG WR WZR ZLQJV 3UHYLRXV ZRUN KDV GHPRQVWUDWHG LWV HIIHFWLYHQHVV ZLWKLQ D SODQDU WZR GHJUHH RI IUHHGRP V\VWHP ZLWK À[HG 90◦ ZLQJ DQJOHV RI DWWDFNV 7KRXJK WKH WUDQVPLVVLRQ ZDV VKRZQ WR

9

SDVVLYHO\ EDODQFH GUDJ LQGXFHG UROO WRUTXHV D À[HG 90◦ DQJOH RI DWWDFN SUHYHQWV WKH ZLQJV IURP JHQHUDWLQJ OLIW 7KLV SDSHU GHVFULEHV D OLIWJHQHUDWLQJ ):0$9 GHVLJQ LQ WHJUDWLQJ SDVVLYHO\ URWDWLQJ ZLQJV ZLWK WKH 'UDJ 3$5,7\ WUDQVPLVVLRQ 9DULDWLRQ RI ZLQJ DQJOH RI DWWDFN LV DFKLHYHG E\ LQFRUSRUDWLQJ D FRPSOLDQW ¶ZLQJ KLQJH· )LJ F LQWR WKH ZLQJ VLPLODU WR WKH DSSURDFK WDNHQ E\ >@ 7KH FRPSOHWH GHVLJQ GHVFULEHG LQ WKLV SDSHU KDV IRXU GHJUHHV RI IUHHGRP D VLJQLÀFDQW LQFUHDVH LQ G\QDPLF FRPSOH[LW\ RYHU WKH WZR GHJUHH RI IUHHGRP V\VWHP SUHYLRXVO\ GHPRQVWUDWHG LQ >@ $ VLQJOH SRZHU DFWXDWRU DSSOLHV DQ RVFLOODWRU\ IRUFH H[FLWLQJ PRWLRQ LQ DOO IRXU GHJUHHV RI IUHHGRP WKURXJK D YDULHW\ RI DHURG\QDPLF DQG LQHUWLDO HIIHFWV 7KH XQGHUDFWXDWHG ÁDSSLQJ ZLQJ V\VWHP LV VKRZQ WR H[HFXWH VWDEOH TXDOLWDWLYHO\ ELRPLPHWLF OLIWJHQHUDWLQJ ZLQJ WUDMHFWR ULHV LQGLFDWLQJ WKDW WKH 'UDJ 3$5,7\ LV D YLDEOH WUDQVPLV VLRQ GHVLJQ IRU LQVHFWVFDOH ):0$9V $ WKHRUHWLFDO PRGHO RI WKH V\VWHP LV GHYHORSHG WR LQYHVWLJDWH WRUTXH EDODQFLQJ FKDUDFWHULVWLFV LQ VLPXODWLRQ $ FRQWURO ¶8QFXW· WULDO ZLWK D QRPLQDOO\ V\PPHWULF V\VWHP GHPRQVWUDWHV SDVVLYH EDODQFLQJ RI UROO WRUTXHV LPSDUWHG IURP HDFK ZLQJ FRPSHQVDWLQJ IRU IDE ULFDWLRQ YDULDWLRQ ,Q ¶&XW· DQG ¶&XW·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À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¶WLPHDYHUDJHG· ZLQJ FRQWURO VHHNLQJ WR FRQWURO DYHUDJH IRUFHV DQG WRUTXHV E\ DSSO\LQJ NLQHPDWLF ZLQJ WUD MHFWRU\ FRUUHFWLRQV RQ D ORQJ WLPHVFDOH ORQJHU WKDQ D ZLQJ ÁDSSLQJ SHULRG >@ $VVXPLQJ WKDW DFWLYH FRQWURO ZLOO QRW EH DWWHPSWHG DW VKRUW VXEZLQJEHDW WLPHVFDOHV WKH TXHVWLRQ RI WKH LGHDO VKRUW WLPHVFDOH EHKDYLRU RI D ZLQJ ÁDSSLQJ PHFKDQLVP LV UDLVHG 7UDGLWLRQDO NLQHPDWLF FRQWURO DSSURDFKHV WDFLWO\ DVVXPH WKDW ULJLG VSHFLÀFDWLRQ RI ZLQJ WUDMHFWRU\ LV D SUHIHUUHG VKRUW WLPHVFDOH EHKDYLRU +RZHYHU WKH VSHFLÀF ZLQJ WUDMHFWRU\ H[HFXWHG LV QRW IXQ GDPHQWDOO\ LPSRUWDQW WR DQ ):0$9 5DWKHU DQ DFWLYH ÁLJKW FRQWURO V\VWHP IRU D URERWLF ÁLHU XVHV WKH ZLQJV DV D WRRO WR JHQHUDWH GHVLUHG UHDFWLRQ IRUFHV DQG WRUTXHV RQ LWV DLUIUDPH

“roll” “pitch”

“yaw”

)LJ  'HÀQLWLRQ RI UROO SLWFK DQG \DZ LQ WKH ERG\ IUDPH

,GHDOO\ WKH ZLQJV VKRXOG H[HFXWH ZKDWHYHU WUDMHFWRULHV DUH QHFHVVDU\ WR UHDOL]H WKHVH GHVLUHG IRUFHV DQG WRUTXHV $ GULYHWUDLQ WKDW SDVVLYHO\ UHJXODWHV WKHVH IRUFHV DQG WRUTXHV DW D VKRUW WLPHVFDOH PD\ VLPSOLI\ D ORQJHU WLPHVFDOH Á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¶'UDJ 3$5,7\· GULYHWUDLQ DQDO\]HG LQ WKLV SDSHU SDVVLYHO\ EDODQFHV ERG\ UROO WRUTXHV LPSDUWHG E\ HDFK ZLQJ +RZHYHU DFWXDWLRQ RI DQ DFWLYH FRQWURO LQSXW FRXOG ELDV V\VWHP G\QDPLFV VXFK WKDW WKH UROO WRUTXH LPSDUWHG E\ RQH ZLQJ LV SDVVLYHO\ UHJXODWHG WR EH  KLJKHU WKDQ WKDW IURP WKH RWKHU ZLQJ 7KLV ORFDO SDVVLYH UHJXODWLRQ PD\ HQDEOH GLUHFW DFWLYH IRUFH DQG WRUTXH FRQWURO DW ORQJ WLPHVFDOHV VLPSOLI\LQJ WKH FRQWURO SUREOHP IRU PDVVOLPLWHG ÁDSSLQJ ZLQJ DHURPHFKDQLFDO SODWIRUPV 6XFK DFWLYH FRQWURO PHFKDQLVPV DUH WKH VXEMHFW RI IXWXUH ZRUN DQG ZLOO QRW EH GLVFXVVHG LQ GHSWK LQ WKLV SDSHU EXW WKHLU EULHI PHQWLRQ VHUYHV WR PRWLYDWH WKH GHWDLOHG DQDO\VLV RI VLPSOHU 3$5,7\ GULYHWUDLQV ZLWKRXW FRQWURO FDSDELOLW\ 7KH IROORZLQJ VHFWLRQV DQDO\]H D VSHFLÀF ):0$9 V\VWHP LQWUR GXFLQJ SDVVLYHO\ URWDWLQJ ZLQJV WR DQ DFWXDWHG 'UDJ 3$5,7\ WUDQVPLVVLRQ

 7KH VSHFLÀF WUDMHFWRU\ PD\ EH LPSRUWDQW IRU HIÀFLHQF\ FRQFHUQV EXW LV LUUHOHYDQW IRU WKH SXUSRVHV RI VWDELOL]LQJ DQG FRQWUROOLQJ WKH DLUIUDPH

 10

,,, 7 +( 0(&+$1,60

7KH WUDQVPLVVLRQ PHFKDQLVP KDV WZR GHJUHHV RI IUHHGRP UHIHUULQJ WR )LJ E q1 LV DFWXDWHG DQG DOORZV SRZHU WR EH LQMHFWHG LQWR WKH V\VWHP ZKLOH q2 LV SDVVLYHO\ GHWHUPLQHG 7KH GHJUHH RI IUHHGRP q2 FRXSOHV WKH XSVWURNH RI RQH ZLQJ WR WKH GRZQVWURNH RI WKH RWKHU DOORZLQJ WKH PHFKDQLVP WR SDVVLYHO\ PRGXODWH ZLQJ VWURNH YHORFLWLHV WR EDODQFH WKH UROO WRUTXHV LPSDUWHG E\ WKH ZLQJV RQ WKH DLUIUDPH $Q LQYHUWLEOH NLQHPDWLF PDSSLQJ UHODWHV q1 DQG q2 WR φR DQG φL  HLWKHU SDLU RI FRRUGLQDWHV FDQ EH XVHG WR GHVFULEH WKH FRQÀJXUDWLRQ RI WKH WUDQVPLVVLRQ $ PRUH GHWDLOHG GHVFULSWLRQ RI WKLV PHFKDQLVP LV DYDLODEOH LQ >@

$ $FWXDWLRQ 3LH]RHOHFWULF DFWXDWLRQ KDV EHHQ FKRVHQ GXH WR LWV KLJK EDQGZLGWK DQG KLJK SRZHU GHQVLW\ >@ 7KH DFWXDWRU LV D EL PRUSK 3=7 FDQWLOHYHU ZLWK D SHDNWRSHDN DFWXDWLRQ VWURNH RI DSSUR[LPDWHO\ μP 7KH EDVH RI WKH FDQWLOHYHU LV JURXQGHG WR WKH ):0$9 DLUIUDPH ZKLOH WKH RXWSXW LV DIÀ[HG WR WKH WUDQVPLVVLRQ LQSXW )LJ   % 7KH 'UDJ 3$5,7\ WUDQVPLVVLRQ 7KH 'UDJ 3$5,7\ WUDQVPLVVLRQ LV D PLOOLPHWHU VFDOH SODQDU OLQNDJH FRQVWUXFWHG XVLQJ 6PDUW &RPSRVLWH 0LFURVWUXFWXUH 6&0 IDEULFDWLRQ WHFKQLTXHV >@ 8QLGLUHFWLRQDO FDUERQ ÀEHU EHDPV IRUP ULJLG OLQNV ZKLOH UHYROXWH MRLQWV DUH UHDOL]HG E\ SRO\PHU ÁH[XUH LQWHUFRQQHFWV 7KH WUDQVPLVVLRQ KDV D VLQJOH DFWXDWHG LQSXW q1 DQG GXDO RXWSXWV GULYLQJ WKH VWURNH DQJOHV RI HDFK ZLQJ 7KH ULJKW ZLQJ VWURNH DQJOH φR LV LOOXVWUDWHG LQ )LJ F ZKLOH WKH OHIW ZLQJ VWURNH DQJOH φL QRW VKRZQ LV WKH DQDORJRXV DQJOH RQ WKH RSSRVLQJ ZLQJ

& :LQJV :LQJV FRQVLVW RI D μP SRO\HVWHU PHPEUDQH VXSSRUWHG E\ FDUERQ ÀEHU YHQDWLRQ VKRZQ LQ )LJ  )DEULFDWHG ZLQJV KDYH PDVVHV XQGHU PJ DQG DUH HIIHFWLYHO\ ULJLG SODWHV H[KLELWLQJ OLPLWHG GHIRUPDWLRQ ZKLOH ÁDSSLQJ ,Q DQ DSSURDFK SLRQHHUHG E\ >@ HDFK ZLQJ LV DWWDFKHG WR D WUDQVPLVVLRQ RXWSXW LQ VHULHV ZLWK D SRO\PHU ÁH[XUH ¶ZLQJ KLQJH· WKDW DOORZV WKH ULJLG ZLQJ WR SDVVLYHO\ URWDWH DURXQG LWV ORQJLWXGLQDO D[LV VHH )LJ F  &RPSOLDQFH DURXQG WKH ZLQJ KLQJH D[LV DOORZV WKH DQJOH RI DWWDFN RI HDFK ZLQJ WR YDU\ SDVVLYHO\ ZKLOH ÁDSSLQJ

Actuator

Airframe

,9 7 +(25(7,&$/ 6 ,08/$7,21 Wing

Transmission

$ $FWXDWLRQ $FWXDWRU GULYH YROWDJH LV WKH VLQJOH LQSXW WR WKH VLPXODWLRQ PRGHO 8VLQJ UHVXOWV IURP D ODPLQDWH SODWH WKHRU\ DQDO\VLV WKH ÀUVW EHQGLQJ PRGH RI WKH FDQWLOHYHU SRZHU DFWXDWRU KDV EHHQ PRGHOHG DV D JURXQGHG VSULQJ LQ SDUDOOHO ZLWK D YROWDJH SURSRUWLRQDO IRUFH >@ 7KH FDQWLOHYHU EHDP KDV D OLQHDU VSULQJ FRQVWDQW RI P1PP DQG XQGHU D 9 DPSOLWXGH VLQXVRLGDO GULYH VLJQDO WKH DFWXDWRU H[HUWV D P1 DPSOLWXGH VLQXVRLGDO IRUFH 7KH GULYH VLJQDO LV DSSOLHG DW +] QHDU PHFKDQLFDO UHVRQDQFH WR LQFUHDVH VWURNH DPSOLWXGH DQG OLPLW UHDFWLYH SRZHU

(a)

q1 ψL

ψR

% 0HFKDQLFDO PRGHO

(b) q2 Wing Hinge

φR ψR (c)

)LJ  D 'LDJUDP RI WKH ):0$9 GHVLJQ E 7KH IRXU GHJUHHV RI IUHHGRP q1  q2  ψ L  DQG ψ R ZLWK UHVSHFW WR DLUIUDPH JURXQG F $ YLHZ RI WKH VKRXOGHU FODULI\LQJ URWDWLRQ DQJOH ψR  5LJKW ZLQJ VWURNH DQJOH φR FDQ EH GHWHUPLQHG IURP q1 DQG q2  DV FDQ φL RI WKH OHIW ZLQJ QRW VKRZQ 

7KH WUDQVPLVVLRQ PHFKDQLVP DORQJ ZLWK WKH ZLQJ KLQJH KDV EHHQ WUHDWHG XVLQJ D SVHXGR ULJLG ERG\ PRGHO >@ $OO FDUERQ ÀEHU OLQNV DUH DVVXPHG WR EH LQÀQLWHO\ VWLII ZKLOH SRO\PHU ÁH[XUH LQWHUFRQQHFWV KDYH EHHQ PRGHOHG DV SHUIHFW UHYROXWH MRLQWV LQ SDUDOOHO ZLWK OLQHDU WRUVLRQ VSULQJV 6SULQJ FRQVWDQWV IRU WKH WUDQVPLVVLRQ MRLQWV DQG ZLQJ KLQJHV KDYH EHHQ FDOFXODWHG XVLQJ FODVVLFDO EHDP WKHRU\ DQG QR GDPSLQJ RU RWKHU LQWHUQDO ORVV PHFKDQLVPV KDYH EHHQ PRGHOHG 7KH ZLQJV WKHPVHOYHV DUH WKH RQO\ VLJQLÀFDQW LQHUWLDV ZLWKLQ WKH V\VWHP DQG DUH WKH RQO\ LQHUWLDV FRQVLGHUHG LQ WKH PRGHO 7KH PDVV RI WKH 6&0 OLQNDJH PHFKDQLVP LV QHJOHFWHG 7KRXJK WKH SLH]RHOHFWULF DFWXDWRU PDVV LV VLJQLÀFDQW GXH WR WKH ODUJH WUDQVPLVVLRQ UDWLR WKH HIIHFWLYH LQHUWLD RI WKH DFWXDWRU LV QHJOLJLEOH DQG KDV EHHQ RPLWWHG IURP WKH PRGHO 7KH ÀQDO WKHRUHWLFDO V\VWHP KDV IRXU GHJUHHV RI IUHHGRP WZR DUH FRQWDLQHG ZLWKLQ WKH 'UDJ 3$5,7\ WUDQVPLVVLRQ ZKLOH WKH WZR ZLQJV HDFK DGG D GHJUHH RI IUHHGRP IURP WKHLU UHVSHFWLYH ZLQJ KLQJHV 7KH RULHQWDWLRQ RI HDFK ZLQJ FDQ EH IXOO\ GHVFULEHG E\ WKH DQJOH RI WKH FRUUHVSRQGLQJ WUDQVPLVVLRQ

 11

ÁDSSLQJ LQ PLQHUDO RLO >@ &DOFXODWLRQ RI ZLQJ URWDWLRQDO PR PHQWV LPSRUWDQW IRU UHDOL]LQJ SDVVLYH ZLQJ URWDWLRQ UHOLHV RQ DGGLWLRQDO H[SHULPHQWDO ZRUN TXDQWLI\LQJ D QRQGLPHQVLRQDO FHQWHU RI SUHVVXUH ORFDWLRQ dˆcp RI IUXLW Á\ ZLQJV >@ >@ 5RWDWLRQDO GDPSLQJ SURSRUWLRQDO WR ψ˙ 2  WKH VTXDUH RI ZLQJ URWDWLRQDO YHORFLW\ KDV EHHQ PRGHOHG LQ DFFRUGDQFH ZLWK H[SHULPHQWDO DQG WKHRUHWLFDO ZRUN RQ WXPEOLQJ SODWHV >@ 7KH FRPSOHWH DHURG\QDPLF PRGHO FDQ EH GLVWLOOHG LQWR WKH IROORZLQJ IRXU DHURG\QDPLF PRPHQWV DSSOLHG WR HDFK ZLQJ ˙ φ˙ 2 CN (ψ) MN = −Ω1 VJQ(φ) ˙ φ˙ 2 CT (ψ) MT = −Ω1 VJQ(φ)

˙ ψ˙ 2 Crd Mrd = −Ω2 VJQ(ψ)   ˙ φ˙ 2 CN (ψ) Mr = − Ω3 dˆcp (ψ) − Ω4 VJQ(φ)

)LJ  7KH ZLQJ ZLWK PHPEUDQH RXWOLQH LQGLFDWHG IRU WKH 8QFXW &XW DQG &XW WULDOV IURP WRS WR ERWWRP $[LV XQLWV DUH LQ PP )RU LQHUWLDO FRPSRQHQWV LQ 7DEOH , 7KH z DQG x FRRUGLQDWH D[HV FRUUHVSRQG WR KRUL]RQWDO DQG YHUWLFDO LPDJH D[HV UHVSHFWLYHO\ :LQJ 7ULDO Ixx Iyy Izz Ixz Ω1 Ω2 Ω3 Ω4

/HIW $OO        

5LJKW 8QFXW        

5LJKW &XW        

5LJKW &XW        

7$%/( , ,QHUWLDO DQG DHURG\QDPLF SDUDPHWHUV XVHG IRU WKH OHIW DQG ULJKW ZLQJV IRU WKH 8QFXW &XW DQG &XW WULDOV $OO YDOXHV KDYH XQLWV RI PJ·PP2  7KH FRRUGLQDWH IUDPH IRU LQHUWLDO FRPSRQHQWV LV GHVFULEHG LQ )LJXUH 

RXWSXW WKH ¶VWURNH DQJOH· φ DQG WKH GHÁHFWLRQ DQJOH RI WKH ZLQJ KLQJH WKH ¶URWDWLRQ DQJOH· ψ  LOOXVWUDWHG IRU WKH ULJKW ZLQJ LQ )LJ F & $HURG\QDPLF PRGHO $HURG\QDPLF HIIHFWV KDYH EHHQ VLPXODWHG XVLQJ D PRGHO GHULYHG IURP WKH EODGH HOHPHQW PHWKRG DVVXPLQJ D SHUIHFWO\ ULJLG ZLQJ SODQIRUP $V PRGHOHG OLIW DQG GUDJ WRUTXHV DUH SURSRUWLRQDO WR φ˙ 2  WKH VTXDUH RI VWURNH YHORFLW\ $YHUDJHG OLIW DQG GUDJ FRHIÀFLHQWV VWURQJ IXQFWLRQV RI WKH URWDWLRQ DQJOH ψ ZHUH WDNHQ IURP H[SHULPHQWDO GDWD FROOHFWHG IURP G\QDPLFDOO\ VFDOHG PRGHOV RI D IUXLW Á\ 'URVRSKLOD PHODQRJDVWHU ZLQJ

   

,Q WKH SUHYLRXV VHW RI HTXDWLRQV MT DFWV DERXW DQ D[LV SHUSHQGLFXODU WR WKH ZLQJ SODQH DQG LV WKH UHVXOW RI DHUR G\QDPLF IRUFHV DFWLQJ LQ WKH ZLQJ SODQH Mr DQG Mrd DUH WKH URWDWLRQDO DQG URWDWLRQDO GDPSLQJ PRPHQWV UHVSHFWLYHO\ ERWK DFWLQJ RQ WKH ZLQJ DURXQG WKH ZLQJ KLQJH D[LV MN DFWV DERXW DQ D[LV SHUSHQGLFXODU WR ERWK WKH ZLQJ SODQH QRUPDO DQG WKH KLQJH D[LV DQG UHVXOWV IURP DHURG\QDPLF IRUFHV QRUPDO WR WKH ZLQJ 7KH WKUHH DHURG\QDPLF FRHIÀFLHQWV UHODWHG WR WDQJHQWLDO CT  QRUPDO CN  DQG URWDWLRQDO GDPSLQJ Crd DHURG\QDPLF IRUFHV DUH GHVFULEHG LQ >@ 7KH SDUDPHWHUV Ω1  Ω2  Ω3  DQG Ω4 KDYH XQLWV RI PJ·PP2 DQG FDQ EH FDOFXODWHG IURP WKH DLU GHQVLW\ ρ DQG WKH VSHFLÀF ZLQJ PRUSKRORJ\ VHH 7DEOH , IRU FDOFXODWHG YDOXHV  5HODWHG ZRUN KDV SURGXFHG H[WHQVLYH H[SHULPHQWDO GDWD YHULI\LQJ WKDW WKLV DHURG\QDPLF PRGHO DGHTXDWHO\ GHVFULEHV SDVVLYH URWDWLRQ RI D VLQJOH ZLQJ H[HFXWLQJ D SUHGHWHUPLQHG VWURNH DQJOH WUD MHFWRU\ DORQJ ZLWK JHQHUDWHG OLIW IRUFHV >@ 7KLV UHIHUHQFHG ZRUN FRQWDLQV D GHWDLOHG GHVFULSWLRQ RI WKH DHURG\QDPLF PRGHO EULHÁ\ VXPPDUL]HG KHUH ' 0DWKHPDWLFDO IRUPXODWLRQ )RU WKH WKHRUHWLFDO PRGHO WKH IRXU FRRUGLQDWHV VSHFLI\LQJ WKH GHYLFH FRQÀJXUDWLRQ ZHUH WDNHQ WR EH WKH OHIW DQG ULJKW ZLQJ VWURNH DQJOHV φL DQG φR  UHVSHFWLYHO\ DORQJ ZLWK WKH OHIW DQG ULJKW ZLQJ URWDWLRQ DQJOHV ψ L DQG ψ R  UHVSHFWLYHO\  7KHVH IRXU TXDQWLWLHV DQG WKHLU WLPH GHULYDWLYHV φ˙ L  φ˙ R  ψ˙ L  DQG ψ˙ R IRUP WKH IXOO HLJKW HOHPHQW VWDWH YHFWRU RI WKH G\QDPLF V\VWHP 7KH ERG\ LQHUWLD RI D URERWLF PJ ):0$9 LV DVVXPHG WR EH RUGHUV RI PDJQLWXGH ODUJHU WKDQ WKH ZLQJ LQHUWLDV 7KLV DVVXPSWLRQ LV UHSUHVHQWDWLYH RI PDQ\ ELRORJLFDO LQVHFWV WKRXJK VRPH H[FHSWLRQV H[LVW HJ EXWWHUÁLHV  $FFRUGLQJO\ WKH ERG\ IUDPH KDV EHHQ WUHDWHG DV DQ LQHUWLDO UHIHUHQFH IUDPH IRU WKH SXUSRVH RI SUHGLFWLQJ ZLQJ G\QDPLFV 7KLV DVVXPSWLRQ DFFXUDWHO\ UHSUHVHQWV WKH JURXQGHGDLUIUDPH H[SHULPHQW XQ GHUWDNHQ LQ 6HFWLRQ 9, )XUWKHUPRUH WKHRUHWLFDO ZLQJ G\QDPLF SUHGLFWLRQV DUH QRW H[SHFWHG WR EH LPSDFWHG VLJQLÀFDQWO\ E\ WKH QRQLQHUWLDO QDWXUH RI WKH ERG\ IUDPH RI D IUHH Á\LQJ ):0$9 7KH HTXDWLRQV RI PRWLRQ IRU WKH ZLQJV ZHUH GHULYHG IURP DQ (XOHU/DJUDQJH IRUPXODWLRQ DVVXPLQJ D À[HG ERG\ IUDPH

 12

30 Drag PARITy Left Wing Right Wing

20 Roll Torque (mN*mm)

Roll Torque (mN*mm)

20

Baseline Left Wing Right Wing

10

0

−10

Baseline Left Wing Right Wing

10

0

4

5

6

7 8 Time (ms)

9

10

11

−30

Baseline Left Wing Right Wing

10

0

−10

−20

3

Drag PARITy Left Wing Right Wing

20

−10

−20

−30

30 Drag PARITy Left Wing Right Wing

Roll Torque (mN*mm)

30

−20

3

4

5

6

D

7 8 Time (ms)

9

10

−30

11

3

4

5

E

6

7 8 Time (ms)

9

10

11

F

)LJ  7KHRUHWLFDO UROO WRUTXH H[SHULHQFHG E\ WKH DLUIUDPH LQ WKH D 8QFXW E &XW DQG F &XW WULDOV

,Q WKH SUHFHGLQJ HTXDWLRQ I L DQG I R DUH WKH LQHUWLDO WHQVRUV RI HDFK ZLQJ FRQVWDQW LQ WKH ZLQJ IUDPH DQG FDOFXODWHG DERXW DQ RULJLQ GHÀQHG E\ WKH FORVHVW SRLQW RQ WKH ZLQJ KLQJH D[LV WR WKH VKRXOGHU D[LV VHH 7DEOH , IRU FDOFXODWHG YDOXHV  7KH VPDOO VKRXOGHU RIIVHW RI WKH 'UDJ 3$5,7\ WUDQVPLVVLRQ KDV EHHQ QHJOHFWHG WKXV QR WUDQVODWLRQDO NLQHWLF HQHUJ\ WHUPV DUH SUHVHQW 7KH WHUP ω  L LV WKH DQJXODU YHORFLW\ RI WKH OHIW ZLQJ D IXQFWLRQ RI φL  ψ L  φ˙ L  DQG ψ˙ L  $Q DQDORJRXV VWDWHPHQW DSSOLHV WR WKH ULJKW ZLQJ DQJXODU YHORFLW\ ω  R 7KH SRWHQWLDO HQHUJ\ V KDV WKH IROORZLQJ IRUP V =

1 2

9  i=1

1 ki γi2 + ka q12 2

35 30 25 20

Drag PARITy Left Wing Right Wing Baseline Left Wing Right Wing

15 10 5



0

7KH IXOO GHYLFH FRQWDLQV QLQH SRO\PHU ÁH[XUH MRLQWV WZR ZLQJ KLQJHV DORQJ ZLWK VHYHQ LQWHUQDO WR WKH 'UDJ 3$5,7\ WUDQVPLVVLRQ 7KH TXDQWLWLHV γi UHSUHVHQW WKH DQJXODU GHÁHFWLRQ RI HDFK ÁH[XUH MRLQW IXQFWLRQV RI φL  ψ L φR  DQG ψ R  7KH FRQVWDQWV ki UHSUHVHQW WKH OLQHDUL]HG WRUVLRQDO VSULQJ FRQVWDQWV IRU HDFK SRO\PHU ÁH[XUH MRLQW 7KH FRQVWDQW ka LV D OLQHDU VSULQJ FRQVWDQW GHVFULELQJ WKH UHVWRULQJ IRUFH RI WKH DFWXDWRU LQ UHVSRQVH WR LWV OLQHDU GHÁHFWLRQ q1  LWVHOI D IXQFWLRQ RI φL DQG φR  7KH /DJUDQJLDQ L LV GHÀQHG DV L = T − V  DQG WKH HTXD WLRQV RI PRWLRQ DUH GHULYHG IURP WKH (XOHU/DJUDQJH HTXDWLRQV IRU HDFK RI WKH IRXU FRRUGLQDWHV pi ∈ {φL , ψ L , φR , ψ R } d ∂L ∂L − = τi dt ∂ p˙ i ∂pi

PRGHO EXW WKH GHWDLOV KDYH EHHQ RPLWWHG IRU EUHYLW\ 7KH IRXU 2QG RUGHU GLIIHUHQWLDO HTXDWLRQV SURGXFHG IURP  ZHUH H[SUHVVHG DV D ÀUVW RUGHU V\VWHP RI HLJKW FRXSOHG QRQOLQHDU GLIIHUHQWLDO HTXDWLRQV $OO WKHRUHWLFDO UHVXOWV IRU WKH 'UDJ 3$5,7\ GHVLJQ DUH WKH UHVXOW RI QXPHULFDOO\ LQWHJUDWLQJ WKHVH GLIIHUHQWLDO HTXDWLRQV XVLQJ D 5XQJH.XWWD EDVHG PHWKRG DV LPSOHPHQWHG E\ WKH 0$7/$% IXQFWLRQ RGH

Yaw Torque (mN*mm)

6LQFH WKH RQO\ PRGHOHG LQHUWLDV LQ WKH V\VWHP DUH WKRVH RI WKH WZR ZLQJV WKH IRUP RI NLQHWLF HQHUJ\ T LV VWUDLJKWIRUZDUG 1  L T L L 1  R T R R I ω  + I ω   ω  ω  T = 2 2



$FWXDWLRQ IRUFH DV ZHOO DV DHURG\QDPLF WRUTXHV DSSHDU LQ WKH PRGHO DV JHQHUDOL]HG IRUFHV τi  $FWXDWLRQ RFFXUV DORQJ q1 VHH )LJ  DQG DHURG\QDPLF WRUTXHV DUH PRUH QDWXUDOO\ FDOFXODWHG LQ WKH ZLQJ IUDPH VR WKH DSSURSULDWH -DFRELDQV KDYH EHHQ XVHG WR PDS WKHVH IRUFHV RQWR WKH FRQÀJXUDWLRQ YDULDEOHV $OO QHFHVVDU\ -DFRELDQ PDWULFHV DQG SDUWLDO GHULYDWLYHV KDYH EHHQ GHULYHG LQ FORVHG DQDO\WLFDO IRUP IRU XVH ZLWKLQ WKH

−5

3

4

5

6

7 8 Time (ms)

9

10

11

)LJ  @

 13

$ FRQWURO WULDO ZKLFK ZLOO EH FDOOHG WKH ¶8QFXW·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φL DQG φR RVFLOODWH RYHU DSSUR[LPDWHO\ 100◦ ZLWK D URWDWLRQ DQJOHV ψ L DQG ψ R RVFLOODWLQJ EHWZHHQ ±60◦  DSSUR[LPDWHO\ 90◦ RXW RI SKDVH 7KHVH URWDWLRQ DQJOHV FRUUHVSRQG WR DQ DQJOH RI DWWDFN α = 90◦ DW VWURNH H[WHQWV DQG α = 30◦ PLGVWURNH 7KHRUHWLFDO ZLQJ WUDMHFWRULHV DUH SORWWHG LQ )LJV D DQG F

)LJ D  ZKHUH WKH WUDQVPLVVLRQ KDV SDVVLYHO\ FRPSHQVDWHG IRU IDEULFDWLRQ DQG DVVHPEO\ HUURU SUHVHQW LQ WKH QRPLQDOO\ V\PPHWULF VWUXFWXUH

Wing Hinges Actuator 5 mm

)LJ  8SSHU LPDJHV DUH V\QFKURQL]HG IUDPHV IURP WKH WZR FDPHUDV GXULQJ WKH 8QFXW WULDO 7UDFNHG SRLQWV DUH LQGLFDWHG DORQJ ZLWK WKHLU WUDMHFWRULHV RYHU WKH FRXUVH RI WKH YLGHR /RZHU LPDJH LOOXVWUDWHV WHVW VWUXFWXUH

Actuator Wing Wing

Transmission

Wing

Wing

Airframe

)LJ  )URQW YLHZ RI WKH H[SHULPHQWDO GHYLFH

9 3$66,9( %2'< 72548( 5(*8/$7,21 7KH 'UDJ 3$5,7\ GHVLJQ GLVWLQJXLVKHV LWVHOI IURP WKH EDVHOLQH GHVLJQ LQ WKH WKHRUHWLFDO UHDFWLRQ WRUTXHV LPSDUWHG E\ WKH ÁDSSLQJ ZLQJV RQ WKH DLUIUDPH 7KH 'UDJ 3$5,7\ LV GHVLJQHG WR EDODQFH WKH UROO UHDFWLRQ WRUTXHV LPSDUWHG E\ WKH ÁDSSLQJ ZLQJV )LJ  SORWV WKH WKHRUHWLFDO UROO WRUTXHV H[SHULHQFHG E\ WKH ERG\ RI D ):0$9 XVLQJ D 'UDJ 3$5,7\ WUDQVPLVVLRQ FRPSDUHG WR WKDW H[SHULHQFHG E\ D ):0$9 XVLQJ D FRQYHQWLRQDO WUDQVPLVVLRQ ,Q DOO WKUHH WULDOV LW LV DSSDUHQW WKDW WKH 'UDJ 3$5,7\ WUDQVPLVVLRQ KDV VXFFHHGHG LQ EDODQFLQJ WKH UROO WRUTXHV H[SHULHQFHG E\ WKH ERG\ GXH WR HDFK ZLQJ 7KH UHVXOWV DUH HVSHFLDOO\ VWULNLQJ LQ WKH 8QFXW WULDO

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ÁDSSLQJ PRWLRQ 3ULRU WR DFTXLULQJ YLGHR WKH FDPHUDV ZHUH

 14

80

80 Uncut 1−Cut 2−Cut

60

Uncut 1−Cut 2−Cut

60

40

40

φ

20

Angle (deg)

Angle (deg)

L

0

−20

0

−20

−40

ψ

L

−40

−60 −80

L

φ

20

ψ

−60 0

1

2

3

4 5 Time (ms)

6

7

8

−80

9

0

1

2

3

4 5 Time (ms)

D

7

8

9

80 Uncut 1−Cut 2−Cut

60

Uncut 1−Cut 2−Cut

60

R

40

40

φ

R

φ

20

Angle (deg)

Angle (deg)

6

E

80

0

−20

20 0

−20

ψR

−40

ψ

−40

−60 −80

L

R

−60 0

1

2

3

4 5 Time (ms)

6

7

8

−80

9

F

0

1

2

3

4 5 Time (ms)

6

7

8

9

G

)LJ  /HIW ZLQJ WUDMHFWRULHV D SUHGLFWHG E\ WKHRU\ DQG E REVHUYHG H[SHULPHQWDOO\ DORQJ ZLWK F WKHRUHWLFDO DQG G H[SHULPHQWDO ULJKW ZLQJ WUDMHFWRULHV (DFK SORW VKRZV UHVXOWV IURP WKH 8QFXW &XW DQG &XW WULDOV

FDOLEUDWHG XVLQJ URXWLQHV IURP WKH &DO7HFK &DPHUD &DOLEUDWLRQ 7RROER[ IRU 0$7/$% >@ 2QFH FDOLEUDWHG WKH WRROER[ DOORZV UHFRQVWUXFWLRQ RI WKUHH GLPHQVLRQDO FRRUGLQDWHV RI SRLQWV LGHQWLÀHG LQ ERWK FDPHUD YLHZV $ +] 9 SHDN WR SHDN VLQXVRLGDO YROWDJH ZDV DSSOLHG WR WKH SRZHU DFWXDWRU DQG V\QFKURQL]HG KLJK VSHHG YLGHR ZDV DFTXLUHG IURP ERWK YLGHR FDPHUDV DW ISV RU  IUDPHV SHU ZLQJVWURNH SHULRG 6DPSOH VWLOO IUDPHV DUH VKRZQ LQ )LJ  7KUHH HDVLO\ GLVWLQJXLVKHG IHDWXUHV RI WKH ZLQJ YHQDWLRQ SDWWHUQ ZHUH PDQXDOO\ WUDFNHG DFURVV  IUDPHV IRU HDFK WULDO ,GHQWLÀFDWLRQ RI DOO WKUHH SRLQWV LQ WZR FDPHUD YLHZV DOORZV VWHUHRSKRWRJUDPPHWULF UHFRQVWUXFWLRQ RI WKH IXOO ZLQJ RULHQWDWLRQ 7KH VLQXVRLGDO GULYH YROWDJH DSSOLHG WR WKH DFWXDWRU KDV EHHQ UHFRUGHG DQG GLJLWL]HG DW N+] V\QFKURQL]HG ZLWK WKH KLJK VSHHG YLGHR VWUHDP 7KH REVHUYHG VWURNH DQG URWDWLRQ DQJOHV DUH SORWWHG DV D IXQFWLRQ RI WLPH LQ )LJ  DORQJ ZLWK SUHGLFWLRQV SURGXFHG E\ WKH WKHRUHWLFDO PRGHO 7LPH V\QFKURQL]DWLRQ KDV EHHQ DFKLHYHG

E\ DOLJQLQJ WKH WKHRUHWLFDO DQG H[SHULPHQWDO GULYH VLJQDOV RPLWWHG IURP WKH SORWV IRU FODULW\ 7KH IXQFWLRQDO IRUP RI WKH DSSOLHG YROWDJH VLJQDO DV D IXQFWLRQ RI WLPH LQ VHFRQGV LV V (t) = 100V + 100V · sin (110 · 2πt)



)URP )LJ  LW LV LPPHGLDWHO\ DSSDUHQW WKDW WKH WKHRUHWLFDO PRGHO DFFXUDWHO\ FDSWXUHV TXDOLWDWLYH FKDUDFWHULVWLFV RI WKH H[SHULPHQWDO PRGHO ZLWK URWDWLRQ DQJOH ψ H[KLELWLQJ DQ DS SUR[LPDWHO\ 90◦ SKDVH ODJ EHKLQG WKH VWURNH DQJOH φ )XUWKHU PRUH WKH WKHRU\ DOVR DFFXUDWHO\ SUHGLFWV RVFLOODWLRQ DPSOLWXGHV IURP DSSOLHG GULYH YROWDJH DQ DFKLHYHPHQW FRQVLGHULQJ WKH FRPSOH[LW\ RI WKLV QRQOLQHDU G\QDPLF V\VWHP 7KHRUHWLFDOO\ SUHGLFWHG WUHQGV LQ ZLQJ WUDMHFWRULHV DV WKH ULJKW ZLQJ SODQIRUP LV DOWHUHG DUH DSSDUHQW LQ H[SHULPHQWDO GDWD 7KH PRGHO SUHGLFWV D PRQRWRQLF LQFUHDVH LQ φR (t) DP SOLWXGH DV SODQIRUP DUHD LV VXFFHVVLYHO\ UHPRYHG IURP WKH ULJKW ZLQJ FRXSOHG ZLWK DQ DVVRFLDWHG GHFUHDVH LQ WKH DPSOLWXGH RI φL (t) 7KLV WUHQG LV UHÁHFWHG LQ WKH H[SHULPHQWDO GDWD DV WKH 'UDJ 3$5,7\ WUDQVPLVVLRQ SDVVLYHO\ GLYHUWV DGGLWLRQDO SRZHU WR WKH XQGHUSHUIRUPLQJ ULJKW ZLQJ 7KH PRGHO DOVR SUHGLFWV D

 15

VXFFHVVLYH GHFUHDVH LQ WKH DPSOLWXGHV RI ERWK ZLQJ URWDWLRQV ψ L (t) DQG ψ R (t) DV ZLQJ PHPEUDQH LV UHPRYHG 7KLV WUHQG LV DSSDUHQW LQ WKH REVHUYHG WUDMHFWRU\ RI ψ L (t) WKRXJK VRPHZKDW DPELJXRXV LQ WKH REVHUYHG WUDMHFWRU\ RI ψ R (t) $PRQJ IHDWXUHV QRW SUHGLFWHG E\ WKLV VLPXODWLRQ PRGHO DUH WKH VTXDUHZDYH DSSHDUDQFH RI REVHUYHG ZLQJ URWDWLRQV DQG WKH FRPSOH[ QRQVLQXVRLGDO GHWDLOV RI VWURNH DQJOH WUDMHFWRULHV ,Q IXWXUH ZRUN LW LV KRSHG WKDW WKHVH GLVFUHSDQFLHV ZLOO EH UHGXFHG E\ D PRUH GHWDLOHG WKHRUHWLFDO PRGHO LQFOXGLQJ IRU H[ DPSOH PHFKDQLFDO ORVV PHFKDQLVPV DQG QRQOLQHDU GHVFULSWLRQV RI SRO\PHU ÁH[XUHV WR EHWWHU SUHGLFW G\QDPLF FKDUDFWHULVWLFV DW ODUJH MRLQW DQJOHV 7KH WUDQVPLVVLRQ GHVLJQ LWVHOI ZLOO EH UHÀQHG WR OLPLW XQLQWHQGHG DQG GLIÀFXOW WR PRGHO EHKDYLRU )RU H[DPSOH RQH VRXUFH RI HUURU LQ WKLV H[SHULPHQWDO WULDO ZDV RIID[LV WUDQVPLVVLRQ FRPSOLDQFH UHVXOWLQJ LQ PHDVXUDEOH GHYLDWLRQ RI WKH ZLQJV IURP WKHLU PHDQ VWURNH SODQHV 9,, & 21&/86,21 $1' ) 8785( : 25. 7KLV SDSHU KDV SUHVHQWHG IXUWKHU HYLGHQFH VXSSRUWLQJ WKH XWLOLW\ RI SDVVLYH XQGHUDFWXDWHG PHFKDQLVPV LQ ):0$9V 6LJQLÀFDQWO\ H[WHQGLQJ SUHYLRXV ZRUN WKH ORDG EDODQFLQJ 'UDJ 3$5,7\ WUDQVPLVVLRQ KDV EHHQ VKRZQ WR EH FRPSDWLEOH ZLWK ORQJLWXGLQDOO\ FRPSOLDQW ZLQJ KLQJHV DOORZLQJ SDVVLYH YDULDWLRQ RI ZLQJ DQJOH RI DWWDFN 7KH UHVXOWLQJ VLQJO\ DFWXDWHG IRXU GHJUHH RI IUHHGRP V\VWHP KDV EHHQ VKRZQ WR H[HFXWH VWDEOH TXDOLWDWLYHO\ ELRPLPHWLF Á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ÀFDWLRQ RI V\VWHP G\QDPLFV )RU H[DPSOH WKH 'UDJ 3$5,7\ WUDQVPLVVLRQ GHVFULEHG LQ WKLV SDSHU H[KLELWV VKRUW WLPHVFDOH G\QDPLFV WKDW EDODQFH UROO WRUTXHV IURP HDFK ZLQJ $Q DFWLYH FRQWURO LQSXW FRXOG EH LQWURGXFHG WR ELDV WKHVH G\QDPLFV VXFK WKDW WKH\ R L DQG τUROO IURP SDVVLYHO\ UHJXODWH WKH UDWLR RI UROO WRUTXHV τUROO WKH OHIW DQG ULJKW ZLQJV UHVSHFWLYHO\ WR D VSHFLÀHG VHWSRLQW q3  L R τUROO /τUROO = q3  1RWH WKDW q3 LV À[HG DW XQLW\ IRU WKH VLPSOH 'UDJ 3$5,7\ WUDQVPLVVLRQ $ YDULHW\ RI G\QDPLF SDUDPHWHUV ZLWKLQ WKH WUDQVPLVVLRQ VXFK DV VSULQJ FRQVWDQWV DQG OLQN OHQJWKV FDQ EH DFWLYHO\ PRGXODWHG DW ORQJ WLPHVFDOHV WR UHDOL]H WKLV ELDVHG VKRUW WLPHVFDOH EHKDYLRU 7KH SRWHQWLDO IRU VLPSOH FRQWURO UHOD WLRQVKLSV VXFK DV   E\SDVVLQJ ZLQJ NLQHPDWLFV WR GLUHFWO\

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ÁLJKW FRQWURO RI LQVHFWVFDOH URERWLF ):0$9V ,W LV KRSHG WKDW IXWXUH UHVHDUFK LQWR WKLV QRYHO PHWKRGRORJ\ ZLOO SURYLGH WRROV WR LQFUHDVH DHURG\QDPLF SHUIRUPDQFH DQG UHGXFH UHTXLVLWH V\VWHP FRPSOH[LW\ KDVWHQLQJ WKH DUULYDO RI DQ DXWRQRPRXV PJVFDOH URERWLF ):0$9 $&.12:/('*(0(17 7KH DXWKRUV JUDWHIXOO\ DFNQRZOHGJH VXSSRUW IURP WKH 1D WLRQDO 6FLHQFH )RXQGDWLRQ $ZDUG 1R &00,   $Q\ RSLQLRQV ÀQGLQJV DQG FRQFOXVLRQV RU UHFRPPHQGDWLRQV H[SUHVVHG LQ WKLV PDWHULDO DUH WKRVH RI WKH DXWKRUV DQG GR QRW QHFHVVDULO\ UHÁHFW WKRVH RI WKH 1DWLRQDO 6FLHQFH )RXQGDWLRQ 5 ()(5(1&(6 >@ 5 - :RRG ´7KH ÀUVW WDNHRII RI D ELRORJLFDOO\ LQVSLUHG DWVFDOH URERWLF LQVHFWµ ,((( 7UDQV 5RE YRO  SS ²  >@ 0 + 'LFNLQVRQ ) 2 /HKPDQQ DQG 6 3 6DQH ´:LQJ URWDWLRQ DQG WKH DHURG\QDPLF EDVLV RI LQVHFW ÁLJKWµ 6FLHQFH YRO  SS ²  >@ : % 'LFNVRQ $ ' 6WUDZ & 3RHOPD DQG 0 + 'LFNLQVRQ ´$Q LQWHJUDWLYH PRGHO RI LQVHFW ÁLJKW FRQWUROµ LQ 3URF $,$$ $HURVSDFH 6FLHQFHV 0HHWLQJ DQG ([KLELW 5HQR 19 -DQXDU\  >@ 5 6 )HDULQJ . + &KLDQJ 0 + 'LFNLQVRQ ' / 3LFN 0 6LWWL DQG - @ 3 6 6UHHWKDUDQ DQG 5 - :RRG ´3DVVLYH DHURG\QDPLF GUDJ EDODQFLQJ LQ D ÁDSSLQJ ZLQJ URERWLF LQVHFWµ - 0HFK 'HV YRO   >@ ; 'HQJ / 6FKHQDWR DQG 6 6DVWU\ ´)ODSSLQJ ÁLJKW IRU ELRPLPHWLF URERWLF LQVHFWV 3DUW LLÁLJKW FRQWURO GHVLJQµ ,((( 7UDQV 5RE YRO  QR  SS ²  >@ 5 - :RRG ( 6WHOW] DQG 5 6 )HDULQJ ´2SWLPDO HQHUJ\ GHQVLW\ SLH]RHOHFWULF EHQGLQJ DFWXDWRUVµ 6HQVRUV $FWXDWRUV $ 3K\VLFDO YRO  QR  SS ²  >@ 5 - :RRG 6 $YDGKDQXOD 5 6DKDL ( 6WHOW] DQG 5 6 )HDULQJ ´0LFURURERW GHVLJQ XVLQJ ÀEHU UHLQIRUFHG FRPSRVLWHVµ - 0HFK 'HV YRO   >@ / / +RZHOO &RPSOLDQW 0HFKDQLVPV -RKQ :LOH\ DQG 6RQV ,QF  >@ - 3 :KLWQH\ DQG 5 - :RRG ´$HURPHFKDQLFV RI SDVVLYH URWDWLRQ LQ ÁDSSLQJ ÁLJKWµ - )OXLG 0HFK YRO ,Q 3UHVV  >@ $ $QGHUVHQ 8 3HVDYHQWR DQG = :DQJ ´8QVWHDG\ DHURG\QDPLFV RI ÁXWWHULQJ DQG WXPEOLQJ SODWHVµ - )OXLG 0HFK YRO  SS ²  >@ - k then Reject observation as an outlier ˜ (b, μ, Σ) 6: return N

4: 5:

7: 8: 9: 10: 11: 12: 13:

K ← ΣA⊤ (AΣA⊤ + R)−1 ˜ ← μ + K(zB − Aμ) μ ˜ ← (I − KA)Σ Σ ¯ ˜ B b+ ← B + diag(μ) μ+ ← 0 ˜ Σ+ ← Σ

where each μci describes the mean offset of a centerline control point, and each μhi describes the mean half-width of the lane estimate at the control point. It is sometimes useful to change the basis curve upon which a lane distribution has been defined, while incurring minimal changes to the actual distribution. Choosing a new mean and covariance in the cases of re-sampled and offset basis curves follows the same procedure as in Sec. IV-B, with minor modifications. When the new basis curve is a variation of the original basis curve, the width components of the mean lane do not change. When the new basis curve is a re-sampling of the original basis curve, the re-sampling matrix H must account for re-sampling the width components in addition to the centerline offset values. B. Observation model, data association, and update A full boundary observation of f is a curve, which we describe with the matrix of control points Z = (z1 , z2 , . . . , zn )⊤ , where each zi can be written: B ¯i + afhi + vi )b zi = bi + (fci (17) ¯i = bi + ziB b

˜ (b+ , μ+ , Σ+ ) return N

VI. L ANE ESTIMATION The boundary curves of a single lane are highly correlated, and information about one boundary gives valuable information about the other. We represent a lane as a piecewise linear centerline curve whose width varies along the curve, and describe its control points with the matrix F = (f 1 , f 2 , . . . , f n )⊤ , where each f i is defined as f i = (fxi , fyi , fhi )⊤ . Using the convention that the normal vectors of a curve point “left”, two points f li and f ri on the left and right boundaries, respectively, can be described as:

fxi + fhi f¯xi fxi − fhi f¯xi f li = f ri = (14) fyi + fhi f¯yi fyi − fhi f¯yi where ¯f i = (f¯xi , f¯yi )⊤ is the normal vector to the centerline curve at point i. A. Lane distributions As with zero-width curves, a basis curve can be used to represent and approximate lanes. We describe the projection f B of f onto b as: B B B , fh1 , fc2 , fh2 , . . . , fcn , fhn )⊤ f B = (fc1

A normal distribution over the projections of f onto b defines a distribution over lanes. We use such a distribution, parameterized by μ and Σ, to represent a belief over the true geometry of f . The mean estimate can be represented by a ˆ = (ˆf 1 , ˆf 2 , . . . , ˆf n )⊤ , where each matrix of control points F ˆ ˆ ˆ control point f i = (fxi , fyi , fˆhi )⊤ can be expressed as: ⎡ˆ ⎤ ⎡ ⎤ fxi bxi + μci¯bxi ˆf i = ⎣fˆyi ⎦ = ⎣byi + μci¯byi ⎦ (16) ˆ μ hi fhi

(15)

B B where fc1 , . . . , fcn is the projection of the centerline of f onto b. Thus, the projection of a lane onto b is simply the projection of its centerline augmented by its half-width terms.

where a has value +1 or −1 for an observation of the left or right boundary, respectively, and we model  the noise terms v = (v1 , v2 , . . . , vn )⊤ jointly as v ∼ N 0, R . Collectively, the offset vector zB = (z1B , z2B , . . . , znB )⊤ can then be expressed as: zB = Af B + v

(18)

where the elements of the observation matrix A are chosen to satisfy Eq. (17). If z is a partial observation of the boundary, such that it projects onto only m control points of b, then A has size 2m × 2n, similar to the case for zero-width curves. Data association and update steps can be approached in the same way as for zero-width curves. Given a lane distribution and observation as expressed above, we can apply a χ2 test to determine if z is an observation of f . When estimating multiple lanes, we use a gated greedy assignment procedure to assign observations to lanes. Once an observation has been associated with a lane estimate, the standard Kalman update steps are used to update the mean and covariance. After the updated estimates have been computed, we once again reparameterize the distribution such that the basis curve coincides with the updated maximum likelihood estimate, to minimize approximation error in future update steps. Fig. 5 shows a full update cycle, where an observation of a lane boundary is used to both update and extend the lane.

 29

(a) A lane estimate

(b) A boundary observation

(c) The updated estimate Fig. 5. A boundary observation is used to update the lane estimate. The middle (black) curve marks the lane centerline, and the outer (blue) curves mark the left and right boundary marginal distributions. Short line segments along the curves mark control points, and the length of these segments indicate 1-σ uncertainty. Note that both boundaries are updated even though only one boundary is observed.

C. Initial estimate We initialize a lane estimate by independently estimating many zero-width curves as described in Sec. V, while periodically searching for curve pairs that are sufficiently long, parallel, and separated by an appropriate distance. Once a suitable pair of boundary curves is identified, they are used to initialize a lane estimate. The initial lane basis curve b is chosen by projecting one boundary curve onto the other and scaling the offset vector by 0.5 (an approximation of the medial axis), and both curve estimates are then reparameterized with b. ˜ (b, μ , Σl ) Referring to the left and right curve estimates as N l ˜ (b, μr , Σr ), we treat them as independent observations and N of the same lane, and express them jointly as: μl Al B (19) z= = f + v = Az f B + v Ar μr where Al and Ar are the transformation matrices relating a lane to its left and right boundary observations (Sec. VI-B), of the unobserved true lane onto b, and f B is the  projection  v ∼ N 0, Σz is a noise term described by:

Σl 0 (20) Σz = 0 Σr Using the information filter [14], we can see that the initial distribution parameters best representing the information provided by the boundary curves can be expressed as: Σ0 μ0

−1 −1 = (A⊤ z Σz Az ) ⊤ −1 = ΣAz Σz z

(21)

VII. E XPERIMENTS To quantitatively assess the performance of our system, we evaluated it against ground truth across two datasets containing data from a forward-facing wide-angle camera (Point Grey Firefly MV, 752x480 @22.8 Hz), and a Velodyne HDL-64E laser range scanner. As input to our system, we used visionand LIDAR-based road paint and curb detection algorithms described in previous work [7].

The first dataset consists of 30.2 km of travel in 182 minutes, and can be characterized by wide suburban lanes, no pedestrians, generally well-marked roads, few vehicles, and a bright early morning sun. The vehicle also traverses a short 0.4 km dirt road and a 1.7 km stretch of highway. The second dataset consists of 13.6 km of travel in 58 minutes through a densely populated city during afternoon rush hour. This dataset can be characterized by roads of varying quality, large numbers of parked and moving vehicles, and many pedestrians. To produce ground truth, we annotated high-resolution georegistered ortho-rectified aerial imagery with lane geometry. The vehicle’s GPS estimates during a data collection provide an initial guess as to the vehicle’s pose; these were corrected by manually aligning sensor data (i.e., camera and LIDAR data) with the aerial imagery at various points in the data collection. The result is a dataset containing ground truth lane geometry relative to the vehicle at every moment of travel. We emphasize that our algorithm uses only local sensor data – GPS and the ground truth map were used only for evaluation purposes. We compare the results of our algorithm, which we refer to as the basis curve (BasCurv) algorithm, with our previous work in the DARPA Urban Challenge [7], which we refer to as the evidence image (EvImg) algorithm. The evidence image algorithm can be used as a standalone lane estimation system by using the output of the first of its two stages, which performs lane detection from sensor data only. Both algorithms use the same features as input. For computational speed, our implementation of the basis curve algorithm used diagonal covariance matrices when estimating lane boundaries, and block-diagonal covariance matrices (2 × 2 blocks) for lane estimation. This introduces additional approximation errors, but yielded good performance in our experiments. After each observation update, basis curves are re-sampled to maintain a uniform (1 m) control point spacing. Parameters such as covariances and data association thresholds were determined experimentally. The basis curve algorithm was implemented in Java and runs at real-time speeds. A. Centerline error The centerline error of a lane estimate at a given point on the estimate is defined as the shortest distance from the estimated lane centerline point to the true centerline of the nearest lane. Fig. 6 shows the 50th and 90th percentile values for the centerline error of the two algorithms as a function of distance from the vehicle. The basis curve algorithm has significantly lower error at all distances. Fig. 7 shows the centerline error as a function of true lane width. The evidence image algorithm assumes a fixed lane width of 3.66 m, and its performance degrades as the true lane width departs from this assumed value. Since the basis curve algorithm jointly estimates lane width and centerline geometry, it remains relatively invariant to changes in lane width.

 30

/RRNDKHDG'LVWDQFH )UDFWLRQRIWUDYHOZLWK ORRNDKHDGGLVWDQFH![

%DV&XUY 

(Y,PJ

/DWHUDOHUURUFP

 

 

 %DV&XUY (Y,PJ

    









Fig. 8.



/RRNDKHDG7LPH

 )UDFWLRQRIWUDYHOZLWK ORRNDKHDGWLPH![

&HQWHUOLQHHUURUYVGLVWDQFHIURPYHKLFOH 

   /RRNDKHDGGLVWDQFHP



%DV&XUY (Y,PJ

    



   /RRNDKHDGWLPHV

Lane estimate lookahead distance (left) and time (right) CDFs.

  







  'LVWDQFHIURPYHKLFOHP



the lane estimates. In (e-g), our method performs well despite adverse lighting scenarios and a snow-covered environment.



Fig. 6. The 50 and 90 percentile values for centerline lateral error, as a function of increasing distance from the vehicle. 0HGLDQFHQWHUOLQHODWHUDOHUURUYVWUXHODQHZLGWK 

%DV&XUY (Y,PJ

/DWHUDOHUURUFP

     



Fig. 7.



  7UXHODQHZLGWKP





Median centerline lateral error as a function of true lane width.

B. Lookahead distance and time The lookahead distance and lookahead time metrics measure how much farther the vehicle can travel before reaching the end of its current lane estimate, assuming constant speed. Lookahead distance is computed by measuring the distance from the vehicle to the farthest point ahead on the current lane estimate, and lookahead time is computed by dividing the lookahead distance by the vehicle’s instantaneous speed. Fig. 8 aggregates lookahead statistics over both datasets, and shows the lookahead distance and lookahead time cumulative distributions for the two algorithms. In all cases, the basis curve algorithm outperforms the evidence image algorithm. For example, the basis curve algorithm provided some lane estimate forward of the vehicle for 71% of distance traveled, compared to 36% for the evidence image algorithm. C. Qualitative results Fig. 9 shows the output of the basis curve lane estimation algorithm in a variety of challenging scenarios. In (a) and (b), tree shadows and an erroneous curb detection are detected and rejected as outliers, leaving the lane estimates intact. In (b), the median strip separating opposite lanes of traffic is correctly excluded from the lane estimates. In (c) and (d), correctly detected road paint is successfully excluded from

VIII. D ISCUSSION Overall, the basis curve algorithm provides lane estimates of equal or better accuracy to those produced by the evidence image approach, and does so more often and with a greater lookahead. We attribute this to the data association and outlier rejection properties of the basis curve algorithm, and to the joint estimation of lane width and centerline geometry. The evidence image algorithm performs no outlier rejection, attempting to fit lanes to both true lane boundary detections and false detections such as tree shadows and non-boundary road paint; nor does it estimate lane width. We have formulated the lane estimation problem in such a way that standard estimation and tracking algorithms based on the Kalman filter can be used for complex lane geometries. In doing so, we gain the advantages of the Kalman filter, and also invite all of its shortcomings. Cases where outliers appear very similar to inliers, such as long shadows nearly parallel to the road, can cause the lane estimates to diverge. Another failure mode arises when one road marking appears initially to be the correct boundary, but the true lane boundary then comes into view. In this case, our method will converge upon the first marking as the lane boundary, since the unimodal nature of the Kalman filter will prevent it from assigning substantial weight to the true boundary. These difficulties are similar to those studied in other estimation domains, and it should also be possible to apply lessons learned in those domains to lane estimation with basis curves. One approach is particle filtering, which has been successfully applied in many estimation and tracking problems to model complex distributions and enable multihypothesis tracking. However, the high dimensionality of the lane estimates would require careful treatment. Finally, in using a Gaussian noise model, we are simplifying and approximating the true system errors. This has proved successful in practice, although more careful study is required to understand the extent to which our simplifications and approximations are valid, and when other models may be more appropriate. IX. C ONCLUSION This paper introduced the notion of basis curves for curve estimation, and described an application to the lane estimation

 31

a.

b.

c.

d.

e.

f.

g. Fig. 9. Lane estimation in a variety of environments. Column 1: Camera images. Column 2: Detections of road paint (magenta) and curbs (green). Column 3: Lane centerline estimates (black) and boundary curve estimates (blue) projected into the image. Column 4: Synthesized overhead view of lane estimates.

problem. A detailed evaluation of our method’s performance on a real-world dataset, and a quantitative comparison against ground truth and a previous approach, shows distinct advantages of the basis curve algorithm, particularly for estimating lanes using partial observations, for handling noisy data with high false-positive rates, and for jointly estimating centerline geometry and lane width. R EFERENCES [1] N. Apostoloff and A. Zelinsky. Vision in and out of vehicles: Integrated driver and road scene monitoring. Int. Journal of Robotics Research, 23(4-5):513–538, Apr. 2004. [2] Y. Bar-Shalom and X.-R. Li. Estimation with Applications to Tracking and Navigation. John Wiley & Sons, Inc., 2001. [3] M. Bertozzi and A. Broggi. GOLD: a parallel real-time stereo vision system for generic obstacle and lane detection. IEEE Transactions on Image Processing, 7(1):62–80, Jan. 1998. [4] M. Bertozzi, A. Broggi, and A. Fascioli. Vision-based intelligent vehicles: State of the art and perspectives. Robotics and Autonomous Systems, 1:1–16, 2000. [5] A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998. [6] E. Dickmanns and B. Mysliwetz. Recursive 3-D road and ego-state recognition. IEEE Trans. Pattern Analysis and Machine Intelligence, 14(2):199–213, Feb. 1992.

[7] A. S. Huang, D. Moore, M. Antone, E. Olson, and S. Teller. Finding multiple lanes in urban road networks with vision and lidar. Autonomous Robots, 26(2-3):103–122, Apr. 2009. [8] A. S. Huang and S. Teller. Lane boundary and curb estimation with lateral uncertainties. In Proc. IEEE Int. Conf. on Intelligent Robots and Systems, St. Louis, Missouri, Oct. 2009. [9] Z. Kim. Robust lane detection and tracking in challenging scenarios. IEEE Trans. Intelligent Transportation Systems, 9(1):16–26, Mar. 2008. [10] Y. Matsushita and J. Miura. On-line road boundary modeling with multiple sensory features, flexible road model, and particle filter. In Proc. European Conference on Mobile Robots, Sep. 2009. [11] J. C. McCall and M. M. Trivedi. Video-based lane estimation and tracking for driver assistance: Survey, system, and evaluation. IEEE Transactions on Intelligent Transport Systems, 7(1):20– 37, Mar. 2006. [12] J. Neira and J. D. Tardos. Data association in stochastic mapping using the joint compatibility test. IEEE Trans. Robotics and Automation, 17(6):890–897, Dec 2001. [13] C. Thorpe, M. Hebert, T. Kanade, and S. Shafer. Vision and navigation for the Carnegie-Mellon Navlab. IEEE Transactions on Pattern Analysis and Machine Intelligence, 10(3):362–373, May 1988. [14] S. Thrun, W. Burgard, and D. Fox. Probabilistic Robotics. MIT Press, 2005. [15] Y. Wang, E. K. Teoh, and D. Shen. Lane detection and tracking using B-Snake. Image and Vision Computing, 22(4):269 – 280, 2004.

 32

Reinforcement Learning to Adjust Robot Movements to New Situations Jens Kober

Erhan Oztop

Jan Peters

MPI for Biol. Cybernetics, Germany Email: [email protected]

ATR Comput. Neuroscience Labs, Japan Email: [email protected]

MPI for Biol. Cybernetics, Germany Email: [email protected]

Abstract—Many complex robot motor skills can be represented using elementary movements, and there exist efficient techniques for learning parametrized motor plans using demonstrations and self-improvement. However, in many cases, the robot currently needs to learn a new elementary movement even if a parametrized motor plan exists that covers a similar, related situation. Clearly, a method is needed that modulates the elementary movement through the meta-parameters of its representation. In this paper, we show how to learn such mappings from circumstances to meta-parameters using reinforcement learning. We introduce an appropriate reinforcement learning algorithm based on a kernelized version of the reward-weighted regression. We compare this algorithm to several previous methods on a toy example and show that it performs well in comparison to standard algorithms. Subsequently, we show two robot applications of the presented setup; i.e., the generalization of throwing movements in darts, and of hitting movements in table tennis. We show that both tasks can be learned successfully using simulated and real robots.

I. I NTRODUCTION In robot learning, motor primitives based on dynamical systems [1], [2] allow acquiring new behaviors quickly and reliably both by imitation and reinforcement learning. Resulting successes have shown that it is possible to rapidly learn motor primitives for complex behaviors such as tennis-like swings [1], T-ball batting [3], drumming [4], biped locomotion [5], ball-in-a-cup [6], and even in tasks with potential industrial applications [7]. The dynamical system motor primitives [1] can be adapted both spatially and temporally without changing the overall shape of the motion. While the examples are impressive, they do not address how a motor primitive can be generalized to a different behavior by trial and error without re-learning the task. For example, if the string length has been changed in a ball-in-a-cup [6] movement1 , the behavior has to be re-learned by modifying the movements parameters. Given that the behavior will not drastically change due to a string length variation of a few centimeters, it would be better to generalize that learned behavior to the modified task. Such generalization of behaviors can be achieved by adapting the meta-parameters of the movement representation2 . In machine learning, there have been many attempts to use meta-parameters in order to generalize between tasks [8]. 1 In this movement, the system has to jerk a ball into a cup where the ball is connected to the bottom of the cup with a string. 2 Note that the tennis-like swings [1] could only hit a static ball at the end of their trajectory, and T-ball batting [3] was accomplished by changing the policy’s parameters.

Figure 1: This figure illustrates a 2D dart throwing task. The situation, described by the state s corresponds to the relative height. The meta-parameters γ are the velocity and the angle at which the dart leaves the launcher. The policy parameters represent the backward motion and the movement on the arc. The meta-parameter function γ(s), which maps the state to the meta-parameters, is learned. Particularly, in grid-world domains, significant speed-up could be achieved by adjusting policies by modifying their metaparameters (e.g., re-using options with different subgoals) [9]. In robotics, such meta-parameter learning could be particularly helpful due to the complexity of reinforcement learning for complex motor skills with high dimensional states and actions. The cost of experience is high as sample generation is time consuming and often requires human interaction (e.g., in cart-pole, for placing the pole back on the robots hand) or supervision (e.g., for safety during the execution of the trial). Generalizing a teacher’s demonstration or a previously learned policy to new situations may reduce both the complexity of the task and the number of required samples. For example, the overall shape of table tennis forehands are very similar when the swing is adapted to varied trajectories of the incoming ball and a different targets on the opponent’s court. Here, the human player has learned by trial and error how he has to adapt the global parameters of a generic strike to various situations [10]. Hence, a reinforcement learning method for acquiring and refining meta-parameters of pre-structured primitive movements becomes an essential next step, which we will address in this paper. We present current work on automatic meta-parameter acquisition for motor primitives by reinforcement learning. We focus on learning the mapping from situations to meta-

 33

parameters and how to employ these in dynamical systems motor primitives. We extend the motor primitives of [1] with a learned meta-parameter function and re-frame the problem as an episodic reinforcement learning scenario. In order to obtain an algorithm for fast reinforcement learning of metaparameters, we view reinforcement learning as a rewardweighted self-imitation [11], [6]. As it may be hard to realize a parametrized representation for meta-parameter determination, we reformulate the rewardweighted regression [11] in order to obtain a Cost-regularized Kernel Regression (CrKR) that is related to Gaussian process regression [12]. We compare the Cost-regularized Kernel Regression with a traditional policy gradient algorithm [3] and the reward-weighted regression [11] on a toy problem in order to show that it outperforms available previously developed approaches. As complex motor control scenarios, we evaluate the algorithm in the acquisition of flexible motor primitives for dart games such as Around the Clock [13] and for table tennis. II. M ETA -PARAMETER L EARNING FOR M OTOR P RIMITIVES The goal of this paper is to show that elementary movements can be generalized by modifying only the meta-parameters of the primitives using learned mappings. In Section II-A, we first review how a single primitive movement can be represented and learned. We discuss how such meta-parameters may be able to adapt the motor primitive spatially and temporally to the new situation. In order to develop algorithms that learn to automatically adjust such motor primitives, we model meta-parameter self-improvement as an episodic reinforcement learning problem in Section II-B. While this problem could in theory be treated with arbitrary reinforcement learning methods, the availability of few samples suggests that more efficient, task appropriate reinforcement learning approaches are needed. To avoid the limitations of parametric function approximation, we aim for a kernel-based approach. When a movement is generalized, new parameter settings need to be explored. Hence, a predictive distribution over the metaparameters is required to serve as an exploratory policy. These requirements lead to the method which we derive in Section II-C and employ for meta-parameter learning in Section II-D. A. Motor Primitives with Meta-Parameters In this section, we review how the dynamical systems motor primitives [1], [2] can be used for meta-parameter learning. The dynamical system motor primitives [1] are a powerful movement representation that allows ensuring the stability of the movement, choosing between a rhythmic and a discrete movement and is invariant under rescaling of both duration and movement amplitude. These modification parameters can become part of the meta-parameters of the movement. In this paper, we focus on single stroke movements which appear frequently in human motor control [14], [2]. Therefore, we will always focus on the discrete version of the dynamical

system motor primitives in this paper (however, the results may generalize well to rhythmic motor primitives and hybrid settings). We use the most recent formulation of the discrete dynamical systems motor primitives [2] where the phase z of the movement is represented by a single first order system z˙ = −τ αz z.

(1)

x˙ 2 = τ αx (βx (g − x1 ) − x2 ) + τ Af (z) , x˙ 1 = τ x2 .

(2)

This canonical system has the time constant τ = 1/T where T is the duration of the motor primitive and a parameter αz , which is chosen such that z ≈ 0 at T . Subsequently, the internal state x of a second system is chosen such that positions q of all degrees of freedom are given by q = x1 , the velocities by q˙ = τ x2 = x˙ 1 and the accelerations by ¨ = τ x˙ 2 . The learned dynamics of Ijspeert motor primitives q can be expressed in the following form

This set of differential equations has the same time constant τ as the canonical system and parameters αx , βx are set such that the system is critically damped. The goal parameter g, a transformation function f and an amplitude matrix A = diag (a1 , a2 , . . . , aI ), with the amplitude modifier a = [a1 , a2 , . . . , aI ] allow representing complex movements. In [2], the authors use a = g −x01 , with the initial position x01 , which ensures linear scaling. Other choices are possibly better suited for specific tasks, see for example [15]. The transformation function f (z) alters the output of the first system, in Equation (1), so that the second system in Equation (2), can represent complex nonlinear patterns and is given by N f (z) = n=1 ψn (z) θ n z. (3) Here, θn contains the nth adjustable parameter of all degrees of freedom, N is the number of parameters per degree of freedom, and ψn (z) are the corresponding weighting functions [2]. Normalized Gaussian kernels are used as weighting functions given by   2 exp −hn (z − cn )  . (4) ψn =  N 2 exp −h (z − c ) m m m=1

These weighting functions localize the interaction in phase space using the centers cn and widths hn . As z ≈ 0 at T , the influence of the transformation function f (z) in Equation (3) vanishes and the system stays at the goal position g. Note that the degrees of freedom (DoF) are usually all modeled independently in the second system in Equation (2). All DoFs are synchronous as the dynamical systems for all DoFs start at the same time, have the same duration and the shape of the movement is generated using the transformation f (z) in Equation (3), which is learned as a function of the shared canonical system in Equation (1). One of the biggest advantages of this motor primitive framework [1], [2] is that the second system in Equation (2), is linear in the shape parameters θ. Therefore, these parameters can be obtained efficiently, and the resulting framework is wellsuited for imitation [1] and reinforcement learning [6]. The

 34

resulting policy is invariant under transformations of the initial position x01 , the goal g, the amplitude A and the duration T [1]. These four modification parameters can be used as the meta-parameters γ of the movement. Obviously, we can make more use of the motor primitive framework by adjusting the meta-parameters γ depending on the current situation or state s according to a meta-parameter function γ(s). The state s can for example contain the current position, velocity and acceleration of the robot and external objects, as well as the target to be achieved. This paper focuses on learning the metaparameter function γ(s) by episodic reinforcement learning. Illustration of the Learning Problem: As an illustration of the meta-parameter learning problem, we take a 2D dart throwing task with a dart on a launcher which is illustrated in Figure 1 (in Section III-B, we will expand this example to a robot application). Here, the desired skill is to hit a specified point on a wall with a dart. The dart is placed on the launcher and held there by friction. The motor primitive corresponds to the throwing of the dart. When modeling a single dart’s movement with dynamical-systems motor primitives [1], the combination of retracting and throwing motions would be represented by one movement primitive and can be learned by determining the movement parameters θ. These parameters can either be estimated by imitation learning or acquired by reinforcement learning. The dart’s impact position can be adapted to a desired target by changing the velocity and the angle at which the dart leaves the launcher. These variables can be influenced by changing the meta-parameters of the motor primitive such as the final position of the launcher and the duration of the throw. The state consists of the current position of the hand and the desired position on the target. If the thrower is always at the same distance from the wall the two positions can be equivalently expressed as the vertical distance. The meta-parameter function γ(s) maps the state (the relative height) to the meta-parameters γ (the final position g and the duration of the motor primitive T ). The approach presented in this paper is applicable to any movement representation that has meta-parameters, i.e., a small set of parameters that allows to modify the movement. In contrast to [16], [17], [18] our approach does not require explicit (re)planning of the motion. In the next sections, we derive and apply an appropriate reinforcement learning algorithm.

Algorithm 1: Meta-Parameter Learning Preparation steps: Learn one or more motor primitives by imitation and/or reinforcement learning (yields shape parameters θ). Determine initial state s0 , meta-parameters γ0 , and cost C 0 corresponding to the initial motor primitive. Initialize the corresponding matrices S, Γ, C. Choose a kernel k, K. Set a scaling parameter λ. For all iterations j: Determine the state sj specifying the situation. Calculate the meta-parameters γj by: Determine the mean of each meta-parameter i γi (sj ) = k(sj )T (K + λC)−1 Γi , Determine the variance σ 2 (sj ) = k(sj , sj ) − k(sj )T (K + λC)−1 k(sj ), Draw the meta-parameters from a Gaussian distribution γj ∼ N (γ|γ(sj ), σ2 (sj )I). Execute the motor primitive using the new meta-parameters. Calculate the cost cj at the end of the episode. Update S, Γ, C according to the achieved result.

policy gradient approaches and natural gradients3 . Reinforcement learning of the meta-parameter function γ(s) is not straightforward as only few examples can be generated on the real system and trials are often quite expensive. The credit assignment problem is non-trivial as the whole movement is affected by every change in the meta-parameter function. Early attempts using policy gradient approaches resulted in tens of thousands of trials even for simple toy problems, which is not feasible on a real system. Dayan & Hinton [19] showed that an immediate reward can be maximized by instead minimizing the Kullback-Leibler divergence D(π(γ|s)R(s, γ)||π ′ (γ|s)) between the rewardweighted policy π(γ|s) and the new policy π ′ (γ|s). Williams [20] suggested to use a particular policy in this context; i.e., the policy π(γ|s) = N (γ|γ(s), σ2 (s)I), where we have the deterministic mean policy γ(s) = φ(s)T w with basis functions φ(s) and parameters w as well as the variance σ 2 (s) that determines the exploration ǫ ∼ N (0, σ 2 (s)I). The parameters w can then be adapted by reward-weighted regression in an immediate reward [11] or episodic reinforcement learning scenario [6]. The reasoning behind this reward-weighted regression is that the reward can be treated as an improper probability distribution over indicator variables determining whether the action is optimal or not. C. A Task-Appropriate Reinforcement Learning Algorithm

B. Problem Statement: Meta-Parameter Self-Improvement The problem of meta-parameter learning is to find a stochastic policy π(γ|x) = p(γ|s) that maximizes the expected return ˆ ˆ J(π) = p(s) π(γ|s)R(s, γ)dγ ds, (5) S

G

where R(s, γ) denotes all the rewards following the selection of the meta-parameter γ according to a situation described T by state s. The return of an episode is R(s, γ) = T −1 t=0 r t with number of steps T and rewards r t . For a parametrized policy π with parameters w it is natural to first try a policy gradient approach such as finite-difference methods, vanilla

Designing good basis functions is challenging, a nonparametric representation is better suited in this context. There is an intuitive way of turning the reward-weighted regression into a Cost-regularized Kernel Regression. The kernelization of the reward-weighted regression can be done straightforwardly (similar to Section 6.1 of [21] for regular supervised learning). Inserting the reward-weighted regression solution w = (ΦT RΦ + λI)−1 ΦT RΓi , and using the Woodbury formula (ΦT RΦ + λI)ΦT = ΦT R(ΦΦT + λR−1 ), we 3 While we will denote the shape parameters by θ, we denote the parameters of the meta-parameter function by w.

 35

(b) Policy after 2 updates: R=0.1

(c) Policy after 9 updates: R=0.8

(d) Policy after 12 updates: R=0.9 3

2

2

2

2

1 0 −1 −2 −3 −4

1 0 −1 −2 −3

−4

−2

0 state

2

4

−4

meta−parameter

3 meta−parameter

3 meta−parameter

meta−parameter

(a) Intial Policy based on Prior: R=0 3

1 0 −1 −2 −3

−4

−2

mean prediction

0 state

2

−4

4

variance

1 0 −1 −2 −3

−4

training points/cost

−2

0 state

2

4

−4

−4

−2

0 state

2

4

Gaussian process regression

Figure 2: This figure illustrates the meaning of policy improvements with Cost-regularized Kernel Regression. Each sample consists of a state, a meta-parameter and a cost where the cost is indicated the blue error bars. The red line represents the improved mean policy, the dashed green lines indicate the exploration/variance of the new policy. For comparison, the gray lines show standard Gaussian process regression. As the cost of a data point is equivalent to having more noise, pairs of states and meta-parameter with low cost are more likely to be reproduced than others with high costs. transform reward-weighted regression into a Cost-regularized Kernel Regression −1  ¯ i = φ(s)T w = φ(s)T ΦT RΦ + λI ΦT RΓi γ  −1 = φ(s)T ΦT ΦΦT + λR−1 Γi , (6)

where the rows of Φ correspond to the basis functions φ(si ) = Φi of the training examples, Γi is a vector containing the training examples for meta-parameter component γ i , and λ is a ridge factor. Next, we assume that the accumulated rewards Rk are strictly positive Rk > 0 and can be transformed into costs by ck = 1/Rk . Hence, we have a cost matrix C = R−1 = diag(R1−1 , . . . , Rn−1 ) with the cost of all n data points. After replacing k(s) = φ(s)T ΦT and K = ΦΦT , we obtain the Cost-regularized Kernel Regression ¯ i = γ i (s) = k(s)T (K + λC)−1 Γi , γ which gives us a deterministic policy. Here, costs correspond to the uncertainty about the training examples. Thus, a high cost is incurred for being further away from the desired optimal solution at a point. In our formulation, a high cost therefore corresponds to a high uncertainty of the prediction at this point. In order to incorporate exploration, we need to have a stochastic policy and, hence, we need a predictive distribution. This distribution can be obtained by performing the policy update with a Gaussian process regression and we directly see from the kernel ridge regression −1

σ 2 (s) = k(s, s) + λ − k(s)T (K + λC) T

k(s),

where k(s, s) = φ(s) φ(s) is the distance of a point to itself. We call this algorithm Cost-regularized Kernel Regression. The algorithm corresponds to a Gaussian process regression where the costs on the diagonal are input-dependent noise priors. Gaussian processes have been used previously for reinforcement learning [22] in value function based approaches while here we use them to learn the policy. If several sets of meta-parameters have similarly low costs the algorithm’s convergence depends on the order of samples. The cost function should be designed to avoid this behavior and to favor a single set. The exploration has to be restricted to safe meta-parameters.

D. Meta-Parameter Learning by Reinforcement Learning As a result of Section II-C, we have a framework of motor primitives as introduced in Section II-A that we can use for reinforcement learning of meta-parameters as outlined in Section II-B. We have generalized the reward-weighted regression policy update to instead become a Cost-regularized Kernel Regression (CrKR) update where the predictive variance is used for exploration. In Algorithm 1, we show the complete algorithm resulting from these steps. The algorithm receives three inputs, i.e., (i) a motor primitive that has associated meta-parameters γ, (ii) an initial example containing state s0 , meta-parameter γ 0 and cost C 0 , as well as (iii) a scaling parameter λ. The initial motor primitive can be obtained by imitation learning [1] and, subsequently, improved by parametrized reinforcement learning algorithms such as policy gradients [3] or Policy learning by Weighting Exploration with the Returns (PoWER) [6]. The demonstration also yields the initial example needed for meta-parameter learning. While the scaling parameter is an open parameter, it is reasonable to choose it as a fraction of the average cost and the output noise parameter (note that output noise and other possible hyper-parameters of the kernel can also be obtained by approximating the unweighted meta-parameter function). Illustration of the Algorithm: In order to illustrate this algorithm, we will use the example of the 2D dart throwing task introduced in Section II-A. Here, the robot should throw darts accurately while not destroying its mechanics. Hence, the cost corresponds to the error between desired goal and the impact point, as well as the absolute velocity of the endeffector. The initial policy is based on a prior, illustrated in Figure 2(a), that has a variance for initial exploration (it often makes sense to start with a uniform prior). This variance is used to enforce exploration. To throw a dart, we sample the meta-parameters from the policy based on the current state4 . After the trial the cost is determined and, in conjunction with 4 In the dart setting, we could choose the next target and thus employ CrKR as an active learning approach by picking states with large variances. However, often the state is determined by the environment, e.g., the ball trajectory in the table tennis experiment (Section III-C) depends on the opponent.



36

D 9HORFLW\

E 3UHFLVLRQ



F &RPELQHG









   

DYHUDJHFRVW

 DYHUDJHFRVW

DYHUDJHFRVW





 



 



 





 QXPEHURIUROORXWV





)LQLWH'LIIHUHQFH*UDGLHQW









 QXPEHURIUROORXWV

5HZDUGïZHLJKWHG5HJUHVVLRQ











 QXPEHURIUROORXWV





&RVWïUHJXODUL]HG.HUQHO5HJUHVVLRQ

Figure 3: This figure shows the performance of the compared algorithms averaged over 10 complete learning runs. Costregularized Kernel Regression finds solutions with the same final performance two orders of magnitude faster than the finite difference gradient (FD) approach and twice as fast as the reward-weighted regression. At the beginning FD often is highly unstable due to our attempts of keeping the overall learning speed as high as possible to make it a stronger competitor. The lines show the median and error bars indicate standard deviation. The initialization and the initial costs are identical for all approaches. However, the omission of the first twenty rollouts was necessary to cope with the logarithmic rollout axis. the employed meta-parameters, used to update the policy5 . If the cost is large (for example the impact is far from the target), the variance of the policy is large as it may still be improved and therefore needs exploration. Furthermore, the mean of the policy is shifted only slightly towards the observed example as we are uncertain about the optimality of this action. If the cost is small, we know that we are close to an optimal policy and only have to search in a small region around the observed trial. The effects of the cost on the mean and the variance are illustrated in Figure 2(b). Each additional sample refines the policy and the overall performance improves (see Figure 2(c)). If a state is visited several times and different meta-parameters are sampled, the policy update must favor the meta-parameters with lower costs. Algorithm 1 exhibits this behavior as illustrated in Figure 2(d). III. E VALUATION In Section II, we have introduced both a framework for meta-parameter self-improvement as well as an appropriate reinforcement learning algorithm used in this framework. In this section, we will first show that the presented reinforcement learning algorithm yields higher performance than off-the shelf approaches. Hence, we compare it on a simple planar cannon shooting problem [23] with the preceding reward-weighted regression and an off-the-shelf finite difference policy gradient approach. 5 In the dart throwing example we have a correspondence between the state and the outcome similar to a regression problem. However, the mapping between the state and the meta-parameter is not unique. The same height can be achieved by different combinations of velocities and angles. Averaging these combinations is likely to generate inconsistent solutions. The regression must hence favor the meta-parameters with the lower costs. CrKR can be employed as a regularized regression method in this setting. The proposed reinforcement learning method only requires a cost associated with the outcome of the trial. In the table tennis experiment (Section III-C), the state corresponds to the position and velocity of the ball over the net. We only observe the cost related to how well we hit the ball. After a table tennis trial, we do not know which state would have matched the employed meta-parameters, as would be required in a regression setting.

The resulting meta-parameter learning framework can be used in a variety of settings in robotics. We consider two scenarios here, i.e., (i) dart throwing with a simulated robot arm, a real Barrett WAM and the JST-ICORP/SARCOS humanoid robot CBi, and (ii) table tennis with a simulated robot arm and a real Barrett WAM. Some of the real-robot experiments are still partially work in progress. A. Benchmark Comparison In the first task, we only consider a simple simulated planar cannon shooting where we benchmark our Reinforcement Learning by Cost-regularized Kernel Regression approach against a finite difference gradient estimator and the rewardweighted regression. Here, we want to learn an optimal policy for a 2D toy cannon environment similar to [23]. The setup is given as follows: A toy cannon is at a fixed location [0.0, 0.1] m. The meta-parameters are the angle with respect to the ground and the speed of the cannon ball. In this benchmark we do not employ the motor primitives but set the parameters directly. The flight of the canon ball is simulated as ballistic flight of a point mass with Stokes’s drag as wind model. The cannon ball is supposed to hit the ground at a desired distance. The desired distance [1..3] m and the wind speed [0..1] m/s, which is always horizontal, are used as input parameters, the velocities in horizontal and vertical directions are predicted (which influences the angle and the speed of the ball leaving the cannon). Lower speed can be compensated by a larger angle. Thus, there are different possible policies for hitting a target; we intend to learn the one which is optimal for a given cost function. This cost function consists of the sum of the squared distance between the desired and the actual impact point and one hundredth of the squared norm of the velocity at impact of the cannon ball. It corresponds to maximizing the precision while minimizing the employed energy according to the chosen weighting. All approaches performed well in this setting, first driving the position error to zero and, subsequently, optimizing the impact

 37

(a) The dart is placed on the launcher.

(b) The arm moves back.

(c) The arm moves forward on an arc.

(d) The arm stops.

(e) The dart is carried on by its momentum.

(f) The dart hits the board.

Figure 4: This figure shows a dart throw in a physically realistic simulation.

(a) The dart is placed in the hand.

(b) The arm moves back.

(c) The arm moves forward on an arc.

(d) The arm continues moving.

(e) The dart is released and the arm follows through.

(f) The arm stops and the dart hits the board.

Figure 5: This figure shows a dart throw on the real JST-ICORP/SARCOS humanoid robot CBi. velocity. The experiment was initialized with [1, 10] m/s as initial ball velocities and 1 m/s as wind velocity. This setting corresponds to a very high parabola, which is far from optimal. For plots, we evaluate the policy on a test set of 25 uniformly randomly chosen points that remain the same throughout of the experiment and are never used in the learning process but only to generate Figure 3. We compare our novel algorithm to a finite difference policy gradient (FD) method [3] and to the reward-weighted regression (RWR) [11]. The FD method uses a parametric policy that employs radial basis functions in order to represent the policy and adds Gaussian exploration. The learning rate as well as the magnitude of the perturbations were tuned for best performance. We used 51 sets of uniformly perturbed parameters for each update step. The FD algorithm converges after approximately 2000 batch gradient evaluations, which corresponds to 2, 550, 000 shots with the toy cannon. The RWR method uses the same parametric policy as the finite difference gradient method. Exploration is achieved by adding Gaussian noise to the mean policy . All open parameters were tuned for best performance. The RWR algorithm converges after approximately 40, 000 shots with the toy cannon. For the Cost-regularized Kernel Regression (CrKR) the inputs are chosen randomly from a uniform distribution. We use Gaussian kernels and the open parameters were optimized by cross-validation on a small test set prior to the experiment. Each trial is added as a new training point if it landed in the desired distance range. The CrKR algorithm converges after approximately 20, 000 shots with the toy cannon. After convergence, the costs of CrKR are the same as for RWR and slightly lower than those of the FD method. The CrKR method needs two orders of magnitude fewer shots than the FD method. The RWR approach requires twice the shots of CrKR demonstrating that a non-parametric policy, as employed by CrKR, is better adapted to this class of problems than a parametric policy. The squared error between the actual and desired impact is approximately 5 times higher for the

finite difference gradient method, see Figure 3. B. Dart-Throwing Now, we turn towards the complete framework, i.e., we intend to learn the meta-parameters for motor primitives in discrete movements. We compare the Cost-regularized Kernel Regression (CrKR) algorithm to the reward-weighted regression (RWR). As a sufficiently complex scenario, we chose a robot dart throwing task inspired by [23]. However, we take a more complicated scenario and choose dart games such as Around the Clock [13] instead of simple throwing at a fixed location. Hence, it will have an additional parameter in the state depending on the location on the dartboard that should come next in the sequence. The acquisition of a basic motor primitive is achieved using previous work on imitation learning [1]. Only the meta-parameter function is learned using CrKR or RWR. The dart is placed on a launcher attached to the end-effector and held there by stiction. We use the Barrett WAM robot arm in order to achieve the high accelerations needed to overcome the stiction. See Figure 4, for a complete throwing movement. The motor primitive is trained by imitation learning with kinesthetic teach-in. We use the Cartesian coordinates with respect to the center of the dart board as inputs. The parameter for the final position, the duration of the motor primitive and the angle around the vertical axis are the meta-parameters. The popular dart game Around the Clock requires the player to hit the numbers in ascending order, then the bulls-eye. As energy is lost overcoming the stiction of the launching sled, the darts fly lower and we placed the dartboard lower than official rules require. The cost function is the sum of ten times the squared error on impact and the velocity of the motion. After approximately 1000 throws the algorithms have converged but CrKR yields a high performance already much earlier (see Figure 6). We again used a parametric policy with radial basis functions for RWR. Designing a good parametric policy proved very difficult in this setting as is reflected by the poor performance of RWR.

 38



&RVWïUHJXODUL]HG.HUQHO5HJUHVVLRQ 5HZDUGïZHLJKWHG5HJUHVVLRQ

 DYHUDJHFRVWVXFFHVV

DYHUDJHFRVW

    

 6XFFHVV &RVW

  

 



  QXPEHURIUROORXWV





Figure 6: This figure shows the cost function of the dartthrowing task for a whole game Around the Clock in each rollout. The costs are averaged over 10 runs with the errorbars indicating standard deviation. This experiment is also being carried out on two real, physical robots, i.e., a Barrett WAM and the humanoid robot CBi (JST-ICORP/SARCOS). CBi was developed within the framework of the JST-ICORP Computational Brain Project at ATR Computational Neuroscience Labs. The hardware of the robot was developed by the American robotic development company SARCOS. CBi can open and close the fingers which helps for more human-like throwing instead of the launcher employed by the Barrett WAM. See Figure 5 for a throwing movement. Parts of these experiments are still in-progress.



  QXPEHURIUROORXWV





Figure 7: This figure shows the cost function of the table tennis task averaged over 10 runs with the error-bars indicating standard deviation. The red line represents the percentage of successful hits and the blue line the average cost. At the beginning the robot misses the ball 95% of the episodes and on average by 50 cm. At the end of the learning the robot hits almost all balls. to update the policy if the robot has successfully hit the ball. Figure 9 illustrates different positions of the ball the policy is capable of dealing with after the learning. Figure 7 illustrates the costs over all episodes. Preliminary results suggest that the resulting policy performs well both in simulation and for the real system. We are currently in the process of executing this experiment also on the real Barrett WAM. IV. C ONCLUSION & F UTURE W ORK

C. Table Tennis In the second evaluation of the complete framework, we use it for hitting a table tennis ball in the air. The setup consists of a ball gun that serves to the forehand of the robot, a Barrett WAM and a standard sized table. The movement of the robot has three phases. The robot is in a rest posture and starts to swing back when the ball is launched. During this swingback phase, the open parameters for the stroke are predicted. The second phase is the hitting phase which ends with the contact of the ball and racket. In the final phase the robot gradually ends the stroking motion and returns to the rest posture. See Figure 8 for an illustration of a complete episode. The movements in the three phases are represented by motor primitives obtained by imitation learning. The meta-parameters are the joint positions and velocities for all seven degrees of freedom at the end of the second phase (the instant of hitting the ball) and a timing parameter that controls when the swing back phase is transitioning to the hitting phase. We learn these 15 meta-parameters as a function of the ball positions and velocities when it is over the net. We employed a Gaussian kernel and optimized the open parameters according to typical values for the input and output. As cost function we employ the metric distance between the center of the paddle and the center of the ball at the hitting time. The policy is evaluated every 50 episodes with 25 ball launches picked randomly at the beginning of the learning. We initialize the behavior with five successful strokes observed from another player. After initializing the meta-parameter function with only these five initial examples, the robot misses ca. 95% of the balls as shown in Figure 7. Trials are only used



In this paper, we have studied the problem of metaparameter learning for motor primitives. It is an essential step towards applying motor primitives for learning complex motor skills in robotics more flexibly. We have discussed an appropriate reinforcement learning algorithm for mapping situations to meta-parameters. We show that the necessary mapping from situation to meta-parameter can be learned using a Cost-regularized Kernel Regression (CrKR) while the parameters of the motor primitive can still be acquired through traditional approaches. The predictive variance of CrKR is used for exploration in onpolicy meta-parameter reinforcement learning. We compare the resulting algorithm in a toy scenario to a policy gradient algorithm with a well-tuned policy representation and the reward-weighted regression. We show that our CrKR algorithm can significantly outperform these preceding methods. To demonstrate the system in a complex scenario, we have chosen the Around the Clock dart throwing game and table tennis implemented both on simulated and real robots. Adapting movements to situations is also discussed in [16] in a supervised learning setting. Their approach is based on predicting a trajectory from a previously demonstrated set and refining it by motion planning. The authors note that kernel ridge regression performed poorly for the prediction if the new situation is far from previously seen ones as the algorithm yields the global mean. In our approach we employ a cost weighted mean that overcomes this problem. If the situation is far from previously seen ones, large exploration will help to find a solution.

 39

(a) The robot is the rest posture.

(b) The arm swings back.

(c) The arm strikes the ball.

(d) The arm follows through and decelerates.

(e) The arm returns to the rest posture.

Figure 8: This figure shows a table tennis stroke on the real Barrett WAM.

(a) Left.

(b) Half left.

(c) Center high.

(d) Center low.

(e) Right.

Figure 9: This figure shows samples of the learned forehands. Note that this figure only illustrates the learned meta-parameter function in this context but cannot show timing and velocity and it requires a careful observer to note the important configuration differences resulting from the meta-parameters. Future work will require to sequence different motor primitives by a supervisory layer. This supervisory layer would for example in a table tennis task decide between a forehand motor primitive and a backhand motor primitive, the spatial meta-parameter and the timing of the motor primitive would be adapted according to the incoming ball, and the motor primitive would generate the trajectory. This supervisory layer could be learned by an hierarchical reinforcement learning approach [24] (as introduced in the early work by [25]). In this framework, the motor primitives with meta-parameter functions could be seen as robotics counterpart of options [9] or macro-actions [26]. R EFERENCES [1] A. J. Ijspeert, J. Nakanishi, and S. Schaal, “Learning attractor landscapes for learning motor primitives,” in Advances in Neural Information Processing Systems 16, 2003. [2] S. Schaal, P. Mohajerian, and A. J. Ijspeert, “Dynamics systems vs. optimal control — a unifying view,” Progress in Brain Research, vol. 165, no. 1, pp. 425–445, 2007. [3] J. Peters and S. Schaal, “Policy gradient methods for robotics,” in Proc. Int. Conf. Intelligent Robots and Systems, 2006. [4] D. Pongas, A. Billard, and S. Schaal, “Rapid synchronization and accurate phase-locking of rhythmic motor primitives,” in Proc. Int. Conf. Intelligent Robots and Systems, 2005. [5] J. Nakanishi, J. Morimoto, G. Endo, G. Cheng, S. Schaal, and M. Kawato, “Learning from demonstration and adaptation of biped locomotion,” Robotics and Autonomous Systems, vol. 47, no. 2-3, pp. 79–91, 2004. [6] J. Kober and J. Peters, “Policy search for motor primitives in robotics,” in Advances in Neural Information Processing Systems 22, 2009. [7] H. Urbanek, A. Albu-Schäffer, and P. van der Smagt, “Learning from demonstration repetitive movements for autonomous service robotics,” in Proc. Int. Conf. Intelligent Robots and Systems, 2004. [8] R. Caruana, “Multitask learning,” Machine Learning, vol. 28, pp. 41–75, 1997. [9] A. McGovern and A. G. Barto, “Automatic discovery of subgoals in reinforcement learning using diverse density,” in Proc. Int. Conf. Machine Learning, 2001.

[10] K. Mülling, “Motor control and learning in table tennis,” Master’s thesis, University of Tübingen, 2009. [11] J. Peters and S. Schaal, “Reinforcement learning by reward-weighted regression for operational space control,” in Proc. Int. Conf. Machine Learning, 2007. [12] C. E. Rasmussen and C. K. Williams, Gaussian Processes for Machine Learning. MIT Press, 2006. [13] Masters Games Ltd., “The rules of darts,” online http://www.mastersgames.com/rules/darts-rules.htm, July 2010. [14] G. Wulf, Attention and motor skill learning. Champaign, IL: Human Kinetics, 2007. [15] D.-H. Park, H. Hoffmann, P. Pastor, and S. Schaal, “Movement reproduction and obstacle avoidance with dynamic movement primitives and potential fields,” in Proc. Int. Conf. Humanoid Robots, 2008. [16] N. Jetchev and M. Toussaint, “Trajectory prediction: learning to map situations to robot trajectories,” in Proc. Int. Conf. Machine Learning, 2009. [17] D. B. Grimes and R. P. N. Rao, “Learning nonparametric policies by imitation,” in Proc. Int. Conf. Intelligent Robots and System, 2008. [18] D. C. Bentivegna, A. Ude, C. G. Atkeson, and G. Cheng, “Learning to act from observation and practice,” Int. Journal of Humanoid Robotics, vol. 1, no. 4, pp. 585–611, 2004. [19] P. Dayan and G. E. Hinton, “Using expectation-maximization for reinforcement learning,” Neural Computation, vol. 9, no. 2, pp. 271–278, 1997. [20] R. J. Williams, “Simple statistical gradient-following algorithms for connectionist reinforcement learning,” Machine Learning, vol. 8, pp. 229– 256, 1992. [21] C. M. Bishop, Pattern Recognition and Machine Learning. Springer Verlag, 2006. [22] Y. Engel, S. Mannor, and R. Meir, “Reinforcement learning with gaussian processes,” in Proc. Int. Conf. Machine Learning, 2005. [23] G. Lawrence, N. Cowan, and S. Russell, “Efficient gradient estimation for motor control learning,” in Proc. Int. Conf. Uncertainty in Artificial Intelligence, 2003. [24] A. Barto and S. Mahadevan, “Recent advances in hierarchical reinforcement learning,” Discrete Event Dynamic Systems, vol. 13, no. 4, pp. 341 – 379, 2003. [25] M. Huber and R. Grupen, “Learning robot control using control policies as abstract actions,” in NIPS’98 Workshop: Abstraction and Hierarchy in Reinforcement Learning, 1998. [26] A. McGovern, R. S. Sutton, and A. H. Fagg, “Roles of macro-actions in accelerating reinforcement learning,” in Grace Hopper Celebration of Women in Computing, 1997.

 40

Analysis and Control of a Dissipative Spring-Mass Hopper with Torque Actuation M. Mert Ankaralı

Uluc¸ Saranlı

Dept. of Electrical and Electronics Eng., Middle East Technical University, 06531 Ankara, Turkey

Dept. of Computer Engineering, Bilkent University, 06800 Ankara, Turkey

[email protected]

[email protected]

Abstract— It has long been established that simple springmass models can accurately represent the dynamics of legged locomotion. Existing work in this domain, however, almost exclusively focuses on the idealized Spring-Loaded Inverted Pendulum (SLIP) model and neglects passive dissipative effects unavoidable in any physical robot or animal. In this paper, we extend on a recently proposed analytic approximation to the stance trajectories of a dissipative SLIP model to analyze stability properties of a planar hopper with a single rotary actuator at the hip. We first describe how a suitably chosen torque controller can compensate for damping losses, maintaining the same energy level across strides and hence reducing the return map to a single dimension. We then identify and characterize equilibrium points for this return map under a fixed leg placement policy and show that “uncontrolled” asymptotic stability is feasible for this energy-regulated system. Subsequent presentation of simulation evidence establishes that the predictions of this approximate model are consistent with the exact plant model. The paper concludes with the application of our energy-regulation scheme to the design of a task-level gait controller that uses explicit leg placement commands in conjunction with the hip torque.

I. I NTRODUCTION Long term practical utility of mobile robots in unstructured environments critically depends on their locomotory aptitude. In this context, the performance of ground mobility that can ultimately be achieved by legged platforms is superior to any other alternative as evidenced by numerous examples in nature as well as a number of very successful dynamically stable autonomous legged robots that have been built to date [10, 25, 26, 30, 37]. Unfortunately, even on flat ground, legged morphologies do not enjoy the simplicity of models supported by the conveniently constrained and continuous modes of ground interaction observed in wheeled and, to some extent, tracked vehicles. Even the most basic legged behaviors such as walking and running require hybrid dynamic models whose analysis and control involve difficult challenges [14, 20, 23]. In the world of quasi-static locomotion with multi-legged robots, one can recover some of this simplicity through active or structural suppression of second order dynamics [39], but these methods are not directly applicable to dynamically dexterous modes of locomotion such as running. One of the most significant discoveries in this context was most likely the recognition of similar center of mass (COM) movement patterns in running animals of widely different sizes

and morphologies [1, 6, 7, 9, 24]. This led to the development of the simple yet accurate Spring-Loaded Inverted Pendulum (SLIP) model to describe such behaviors [21, 34]. Significant research effort was devoted to both the use of this model as a basis for the design of fast and efficient legged robots [10, 19, 27, 30] as well as its analysis to reveal fundamental aspects of associated locomotory behaviors [20]. The present paper falls into the latter category and contributes by investigating the previously unaddressed question of how the presence of passive damping affects the behavioral characteristics of running with the SLIP model. Our treatment of this question is based on the use of analytic approximations to the otherwise non-integrable stance dynamics of the model. A number of such approximations have already been proposed in the literature. In particular, [35] uses a Hamiltonian formulation of the SLIP dynamics with an iterative application of the mean-value theorem to obtain an accurate return map for symmetric steps. More recently, [16] presents an analytically simple approximation to the stance dynamics of a conservative SLIP equipped with a linear spring. This approach is based on a linearization of gravity around mid-stance, similar in form to the solution proposed in [28] but involving a much more carefully formulated derivation. Inaccuracies of these approximations in the presence of nonsymmetric gravitational effects were partially addressed in [5] using explicit corrections. In contrast to the lossless models adopted in all these approaches, a new return map for the lossy SLIP model with viscous damping in the leg was proposed in [4], providing a basis for our present inquiry. Despite the availability of methods to analyze stability properties of locomotory behaviors in the absence of closed-form expressions for a Poincar´e map [2, 3], a number of different possible approaches become available once a sufficiently accurate analytic return map is available. For example, [18] investigates in depth stability properties of a SLIP model attached to a rigid body by neglecting the effects of gravity, which allows for the derivation of suitable closed-form expressions for stride trajectories. A similar but less rigorous stability analysis was provided in [16] with comparisons to previous numerical results in [36] as well as biological data. In contrast to the lack of feedback control in these characterizations, [33] studies the stability of the SLIP model under a novel leg placement

 41

body reaches its maximum height during flight with y˙ = 0. Another important event, not illustrated in the figure, is bottom, corresponding to the point of maximal leg compression during stance. Table I details the notation used throughout the paper.

descent

Fig. 2.

k y

c

θ

Fig. 1 illustrates the Torque-actuated Dissipative SpringLoaded Inverted Pendulum (TD-SLIP) plant we investigate in this paper. It consists of a fixed orientation (2-DOF) planar rigid body with mass m, connected to a massless, fully passive leg with linear compliance k, rest length r0 and linear viscous damping c, through an actuated rotary joint with torque τ . The system alternates between stance and flight phases during running, with the flight phase further divided into the ascent and descent subphases. Fig. 2 illustrates the three important events that define transitions between these phases: touchdown, where the leg comes into contact with the ground, liftoff, where the toe takes off from the ground and finally apex, where the

apex

liftoff ascent

A single TD-SLIP stride with definitions of transition states TABLE I

System States, Event States and Control Inputs x, y, x, ˙ y˙ Cartesian body position and velocities r, θ, r, ˙ θ˙ Leg length, leg angle and velocities τ Hip torque command during stance ya , x˙ a Apex height and velocity θtd , r˙td , θ˙td Touchdown leg angle, polar velocities tb , rb , θb Bottom time, leg length and angle tlo , rlo , θlo , r˙lo , θ˙lo Liftoff time, leg length, angle and velocities pθ Angular momentum around the toe Kinematic and Dynamic Parameters m, g Body mass and gravitational acceleration k, r0 , c Leg stiffness, rest length and damping

During flight, the body obeys ballistic flight dynamics



x ¨ 0 = y¨ −g

and the massless leg can be arbitrarily positioned. In contrast, during stance, the toe remains stationary on the ground while the body mass feels forces generated by both the passive spring-damper pair and the hip torque. The stance dynamics of the planar SLIP model in polar leg coordinates with respect to the toe location take the form d dt

x Fig. 1. TD-SLIP : Dissipative spring-mass hopper with rotary hip actuation

stance

N OTATION USED THROUGHOUT THE PAPER

A. System Dynamics and the Apex Return Map

r

[rlo , θlo , r˙lo , θ˙lo ]

touchdown

[θtd , r˙td , θ˙td ]

II. T HE T ORQUE -ACTUATED D ISSIPATIVE SLIP M ODEL

τ

[ya , x˙ a ]k+1

[ya , x˙ a ]k

apex

control strategy that only relies on easily obtainable state measurements. Another inquiry on how to achieve effective control of the SLIP model, now extended to a spatial setting, is provided in [8], focusing on lateral motions of the model. In conjunction with these studies primarily focused on running behaviors, similar analytically motivated contributions were also made to the structurally different walking behaviors [15, 22, 40], providing evidence that the same dynamic model can provide a unified description for both walking and running [17]. Our contributions in the present paper have a number of important differences from existing work. Firstly, our plant model is dissipative, impairing the accuracy of most existing analytic approximations and associated predictions. Secondly, in contrast to the usual energy regulation mechanisms in the literature through adjustments of the leg length or changing stiffness, our model uses only a single torque actuator at the hip relative to a virtual body with fixed orientation to compensate for energy losses. These changes are motivated by being much more realistic from an implementation point of view, as evidenced by the successful use of similar actuation mechanisms in the Scout quadrupeds [26] and the RHex hexapod [30] as well as a number of other monopedal platforms [12, 32]. Finally, our approximate solutions to the return map also take into account the effect of gravity on the angular momentum for steps that are non-symmetric with respect to the gravitational vertical, increasing the practical applicability of associated stability results.



mr˙ mr 2 θ˙



=



mrθ˙2 − mg cos θ − k(r − r0 ) − cr˙ mgr sin θ + τ



. (1)

A very useful abstraction for the analysis and control of cyclic TD-SLIP trajectories is provided by the apex return map, defined as a Poincar´e map from one apex point to the next. In the following sections, we will use this map to study stability properties of TD-SLIP, and later adopt it as a tasklevel gait representation for a closed-loop running controller. The apex return map can be formulated as P := Pa ◦Ps ◦Pd by composing three individual submaps Pd , Ps , Pa for the descent, stance and ascent phases, respectively. The descent and ascent maps are trivial and are given by    −x˙ a r˙td  (2) ) = R(π/2 − θ td r0 θ˙td 2g(ya − r0 cos θtd )     2 ya rlo cos θlo + y˙ lo /(2g) = (3) Pa : x˙ a x˙ lo Pd :

 42



where x˙ lo and y˙ lo are liftoff velocities in Cartesian coordinates and R denotes the standard 2D rotation matrix. Unfortunately, the dynamics of (1) are not integrable in closed form. Consequently, we will use an analytical approximation for the stance map, which we describe in the next section.

The final step in completing the stance map requires finding the time of liftoff. Only one of the two liftoff conditions described in [4] is applicable in the context of the present paper since we do not allow control of the liftoff leg length. Consequently, the liftoff time is solely determined by the ˙ lo ) = 0, for solution to the equation k(r0 − r(tlo )) − c r(t which a sufficiently accurate analytical approximation can be found by approximating the exponential coefficient in the radial solution of (6) by its value at a specific instant during decompression. In particular, noting that the compression and decompression times are roughly equal, we use e−ζ ωˆ 0 t ≈ e−ζ ωˆ 0 2tb , where tb denotes the bottom time, easily found by solving (7). Under this assumption, we have

B. An Approximate Stance Map for the Unforced TD-SLIP A new analytical approximation to the dynamics of a dissipative SLIP model was proposed in [4]. However, this method assumes the presence of radial leg actuation, either in the form of a controllable leg stiffness, or the regulation of touchdown and liftoff leg lengths. In this section, we briefly review their method and extend it to support the hip torque actuation of our model. The approximation proposed in [4], which, in turn, is based on the methods described in [16], relies on two key assumptions: 1). The angular travel throughout stance is relatively small and remains close to the vertical, allowing linearization of the gravitational potential in the Lagrangian with subsequent conservation of the angular momentum pθ := mr2 θ˙ and 2). the radial compression is small with r0 − r ≪ r0 , allowing a truncated Taylor expansion of related terms. As described in [4], under these conditions and assuming, for now, that τ = 0, the radial component of (1) reduces to

tlo

−

ω02 , (5) = e−ζ ωˆ 0 t (A cos(ωd t) + B sin(ωd t)) + F/ˆ  where ˆ 0 := ω02 + 3ω 2 , ζ := c/(2mˆ ω0 ), ωd :=  we have ω 2 2 2 ω ˆ 0 1 − ζ , F := −g+r0 ω0 +4r0 ω and A and B determined by touchdown states as

Simple differentiation and further simplification yields radial TD-SLIP trajectories as (6) (7)

with M , φ and φ2 determined through trigonometric identities. At this point, the angular trajectories can be determined using the constant angular momentum. An additional linearization of the term 1/r2 leads to an analytical solution for the rate of change of the leg angle as ˙ θ(t)

=

ˆ 02 ) − 3ω − 2ωF/(r0 ω −ζ ω ˆ0t 2ωM e cos(ωd t + φ)/r0 ,

(8)

integrated to yield the angular trajectory θ(t)

= θtd + X t + −ζ ω ˆ0 t

Y (e

(9) cos(ωd t + φ + φ3 ) − cos(φ + φ3 )).

with X, Y and φ3 computed accordingly as in [4].

(11)

C. Stance Map for the Torque Controlled TD-SLIP

ω02 , A := r0 − F/ˆ B := (r˙td + ζ ω ˆ 0 A)/ωd .

r(t) ˙

⎤ r(tlo ) θ(tlo ) ⎥ ⎥ . r(t ˙ lo ) ⎦ ˙ lo ) θ(t

where the right hand side is a function of touchdown states. Note, however, that these derivations completely ignore the presence of the hip torque. In the next section, we propose a new method to incorporate the effects of the hip torque through a fixed correction on the angular momentum value pθ in a way similar to the one used in [5] for gravity corrections.

r(t)

= M e−ζ ωˆ 0 t cos(ωd t + φ) + F/ˆ ω02 , = −M ω ˆ 0 e−ζ ωˆ 0 t cos(ωd t + φ + φ2 ) ,

(2π − arccos(k(r0 − F/ˆ ω02 )/(M M e−ζ ωˆ 0 γtb )) (10) φ − φ4 )/ωd ,

which yields the stance map as ⎡ ⎤ ⎡ rlo ⎢ θlo ⎥ ⎢ ⎥ ⎢ Ps : ⎢ ⎣ r˙lo ⎦ = ⎣ θ˙lo

r¨ + (c/m)r˙ + (ω02 + 3ω 2 )r = −g + r0 ω02 + 4r0 ω 2 , (4)  where we define ω0 := k/m and ω := pθ /(mr02 ). Solutions to this simple second-order ODE can be found as

r(t)

≈

Hip actuation in legged systems can serve a number of different purposes. Among both biological [1] and robotic [13, 19, 27] systems, its most common uses involve retraction of legs in flight and control of body posture with legs in stance. Interestingly, the use of hip actuation to provide thrust has not been studied as extensively in the robotics literature. In addition to a few direct experimental inquiries [12, 32] and indirect uses in multi-legged platforms [26, 30], it has received limited attention in [2] in the form of an active spring. In the present paper, we propose an open-loop hip actuation regime that enforces the ramp torque profile  τ0 (1 − ttf ) if 0 ≤ t ≤ tf (12) τ (t) = 0 if t > tf during stance, with τ0 and tf chosen prior to touchdown. This open-loop profile has three important advantages. Firstly, its simple functional dependence on time allows us to easily incorporate its effects into the derivations of the previous section. Second, if we choose tf to be the liftoff time, we have τ (tlo ) = 0, which prevents premature leg liftoff due to the action of the hip and ensures a structural match to the trajectories of the unforced system. Finally, its unidirectional action ensures that no negative work is done during stance. Inspection of the TD-SLIP dynamics of (1) shows that the hip torque directly acts on the angular dynamics and only indirectly effects radial motion. Consequently, we hypothesize

 43

that an average correction to the constant angular momentum pθ of Section II-B can capture the effects of the hip torque on system trajectories. Normally, the angular momentum during stance can be formulated as  t  t mgr(η) sin θ(η)dη, (13) τ (η)dη + pθ (t) = pθ (0) + 0

0

by integrating the angular dynamics. Adopting the method proposed in [5], we compute a corrected angular momentum pˆθ = pθ (0) + Δpτ + Δpg .

(14)

where Δpτ and Δpg incorporate the time averaged effects of the leg torque and gravitational acceleration, respectively. Assuming tf = tlo , we have   tlo  η1 tlo 1 Δpτ := . (15) τ (η2 )dη2 dη1 = τ0 tlo 0 3 0 However, even with available analytic approximations, derivation of an exact closed-form expression for Δpg is not feasible. Instead, we use a linear approximation to the integrand r(η) sin θ(η) using its values at the touchdown and liftoff, resulting in mgtlo (16) (2r0 sin θtd + rlo sin θlo ) . 6 Estimated values for the liftoff time tlo , leg angle θlo and leg length rlo are provided by the unforced approximations of the previous section. Substituting pˆθ for the constant angular momentum in all derivations of Section II-B, we obtain a new approximation that takes into account the effects of both the hip torque and gravity on the stance trajectories. Note that the corrections we propose have an iterative character since both (15) and (16) use prior estimates of tlo and θlo . Consequently, starting from the unforced approximations, it is possible to iteratively apply these corrections to obtain more accurate predictions at the expense of analytic simplicity. Our simulations show that more than a single iteration is only needed for extreme conditions such as the angle of attack being very close to the touchdown leg angle, causing a bounce-back. Δpg :=

and keep it constant across subsequent strides. The underlying idea is that since damping losses are proportional to the total energy level of the system, constant energy input will give rise to trajectories at a stable energy level. Unfortunately, in our 2-DOF model, such a strategy results in a two dimensional return map, for which, analytical solution and characterization of equilibrium points is not feasible. Another possibility, which we adopt in the present paper, is to use the hip torque to compensate for all dissipative effects within a single step, ensuring conservation of energy in the apex return map and hence reducing its dimension by one. Note that the total energy dissipated within a single TD-SLIP step is given by (17) Eloss = Ec + Ek , where Ec represents damping losses with  tlo cr˙ 2 (η) dη , Ec :=

and Ek := (rlo − r0 )2 /2 captures the leftover energy in the leg spring when it lifts off before it is fully extended due to damping. Fortunately, our analytic approximations provide closed form expressions for both of these components. In particular, damping losses can be approximately computed as Ec

Unlike previous stability studies of lossless spring-mass hoppers, fully passive self-stabilization with a fixed touchdown angle and no active control is not possible with the TDSLIP model since damping losses will eventually drain out all energy in the system. Consequently, active hip thrust must be employed to sustain locomotion. Recall that our choice of the hip torque in (12) incorporates two parameters: τ0 and tf . We have already shown that choosing tf = tlo is advantageous in preventing early liftoff and ensuring structural correspondence of system trajectories to our analytical approximation. The simplest possible strategy for the remaining parameter τ0 , very close in spirit to the radial actuation strategy adopted by Raibert’s runners [27] and its subsequent analysis in [21], is to choose a particular value

= −

−c/M 2 ω ˆ0 (ζ cos(2(φ + φ2 ) + φ3 ) + 1 (19) 4ζ e−2ζ ωˆ 0 tlo (ζ cos(2ωd tlo + 2(φ + φ2 ) + φ3 ) + 1)) ,

while Ek only depends on the previously computed rlo . In contrast, the energy supplied by the hip torque is  tlo t ˙ (1 − ) θ(t)dt , (20) Eτ = τ0 tlo 0 for which our analytical approximations can also be used to obtain closed-form expressions. We omit the details here for space considerations. Since both (17) and (20) can be obtained in closed form as a function of initial conditions and the choice of touchdown angle θtd , we can easily find the desired torque magnitude τ0 by solving

III. S TABILITY OF AN E NERGY-R EGULATED TD-SLIP A. Compensation of Damping Losses

(18)

0

Eτ

=

Eloss .

(21)

As noted above, this choice of torque results in successive apex states having the same energy, at least while working within our approximate apex return map. Naturally, additional corrections would be needed to apply these ideas to the exact plant model since inaccuracies of our approximations would invalidate this conservation. Nevertheless, we use this active compensation regime to reduce the dimension of our analytic apex return map, allowing us to easily identify its equilibrium points and characterize their stability. B. Equilibrium Points with a Fixed Leg Placement Policy In this section, we use our analytic approximations to identify and characterize equilibrium points of the one dimensional “energy-regulated” return map on the apex height ya arising from the use of a fixed touchdown angle policy with

 44

2

θtd = β and the energy-regulating hip torque described in Section III-A. All results in this section will be presented in non-dimensional versions of relevant variables, defined as

ζ0

2

:= Ea /(mgr0 ) := kr0 /(mg) √ := c/(2 mk) .

4

y¯a

1.6 1.4 1.2 1

1

1.5

2

1.6

1.8

2.5

3

y¯a

2

1

2

3

E¯a

4

3.5

y¯a

3 2.5 2 1.5 1

y¯a

3.5

5

1.5

1

0.5

1

1.8

1.4

4 3

2

1.2

Mean-Square % Error in E¯a

:= ya /r0 √ := x˙ a / gr0

Finally, in order to facilitate comparison with earlier studies, we use kinematic and dynamic parameters that roughly match those of an average human with m = 80kg and r0 = 1m. Fig. 3 shows two families of return maps for β = 20◦ and β = 32◦ , respectively, together with the dependence of equilibrium points on the energy level of the system. These results show that the TD-SLIP exhibits asymptotically stable behavior under the fixed touchdown angle, energy-regulated regime, with the location of the equilibrium point depending on the chosen energy level. We can also observe that as the fixed touchdown angle β increases, the energy range for which stable fixed points exist increases as well. This is rather natural since the torque actuation at the hip can only supply energy through the angular momentum, which directly increases the angular span during stance. Increasing the touchdown angle admits a larger angular span for stance, allowing stable fixed points to form at higher energy levels as well.

1

5

y¯a

y¯a ¯ x˙ a ¯a E k¯

6

10

E¯a

15

Fig. 3. Apex height return map (left) and associated equilibrium points (right) for the TD-SLIP model as a function of different (dimensionless) energy levels, generated with the proposed analytical approximations. The top ¯ = 40, ζ0 = 0.07, whereas the bottom plots are obtained with β = 20◦ , k ¯ = 40, ζ0 = 0.07 in dimensionless coordinates. Solid plots use β = 32◦ , k and dashed lines in the right figure indicate stable and unstable equilibrium points respectively. Shaded regions correspond to kinematically infeasible configurations.

Having established the presence of stable equilibrium points for the torque-controlled SLIP model, Fig. 4 shows a comparison of fixed points predicted by our analytic approximations,

6

E¯a

8

10

0

4

6

E¯a

8

10

Fig. 4. Left: Comparison of stable equilibrium points predicted by our analytic approximation (solid line) with those obtained by numeric simulation of TD-SLIP dynamics (plus signs) for β = 28◦ and different apex energy ¯a ∈ [2, 10]. Shaded region in the middle illustrates the levels in the range E stable domain of attraction for the simulated plant model. Right: Percentage mean-square error between initial and steady-state dimensionless energy levels for the simulated plant.

with those that arise within simulations of the exact TDSLIP model. In order to make direct comparisons possible, we started TD-SLIP simulations from a large range of initial ya and Ea values, with a fixed touchdown angle and an energy regulation controller similar to the one presented Section IIIA, but now taking the energy level of the very first step as an overall regulation goal. This modification was necessary since using the approximations to locally enforce energy conservation at every step would slowly cause prediction errors to accumulate, either draining all energy out of the system, or causing it to diverge. We then checked whether the system converges to a stable equilibrium point in apex coordinates after 25 steps up to a tolerance of 10−4 . Shaded region in the middle of the left plot of Fig. 4 illustrates the resulting domain of attraction, while the plus marks in the same plot illustrate the associated set of fixed points. Note, also, that the domain of attraction exhibited by the simulation almost exactly covers the region between the unstable and stable fixed points predicted by our approximations. There is also an almost exact match between the fixed points predicted by our approximations and those obtained from simulation. The cavities to the right of the region of attraction arise from the presence of the “gap” region in the return map, resulting from kinematic constraints that require the apex height to be sufficiently large to allow leg placement. The reason for this can be clearly seen in the bottom right plot of Fig. 3, where parts of the return map overlap with the kinematically infeasible gray region on the bottom. This means that some initial conditions at high energy levels will lead to apex states for which leg placement at an angle of β is impossible. This gap was also observed by previous studies [16], and is reproduced by both our analytical approximations, and the simulated plant. The right plot in Fig. 4 shows the mean and standard deviations of the percentage mean-square energy difference between the initial and steady state apex points for the simulated plant. The fact that this difference is consistently below 0.3% shows that our approximations are capable of very accurately modeling energy losses and successfully predict fixed points of

 45

the exact TD-SLIP plant. It is worth noting, also, that accuracy also increases significantly with increasing energy levels.

Center of Mass Trajectory

C. Parameter Dependence of Equilibrium Points Equilibrium points that arise from our fixed touchdown angle, energy-regulated regime naturally depend on the kinematic and dynamic parameter choices. Fig. 5 illustrates the dependence of stable fixed points on each individual parameter (the touchdown angle β, the dimensionless leg stiffness k¯ or leg damping ζ0 ) with the remaining two parameters kept constant. The leftmost figure mirrors our observations in the previous section, namely that the range of stable energy levels increase with larger touchdown angles. k¯ = 40, ζ0 = 0.07 2

2

1.8

1.8

1.6

1.6

1.6

1.4

1.4

y¯a

2

β

1.2

1 0.8

1.2

1

5

10

E¯a

15

20

0.8

introduced in the beginning of the stance phase, resulting in associated virtual footfalls appearing behind the actual toe location. Towards the end of the stance phase, the hip torque approaches zero and brings the virtual footfall and actual toe locations together. This qualitative structure is observed for all steady-state trajectories of the TD-SLIP model and is remarkably consistent with biological data presented in [38]. Even though we do not yet have any quantitative basis in which any predictive claims can be made, we think that this correspondence may provide evidence towards the use of hip torque as an additional source of energy used by biological runners, improving the predictive accuracy and utility of dynamic models of running.

1.4

k¯

ζ0

1

5

10

E¯a

15

20

0.8

5

10

E¯a

15

Leg Force Directions

Fig. 6. COM trajectory for a single steady-state stride of the TD-SLIP ¯ = 40, ζ0 = 0.07, running at approximately 3m/s (0.96 in model with k dimensionless units). Comparison of ground reaction force directions during stance to biological data presented in [38] reveals a remarkable qualitative match.

β = 28◦, k¯ = 40

β = 28◦ , ζ0 = 0.07

1.8

1.2

Leg length

20

Fig. 5. Dependence of stable equilibrium points on variations of the ¯ (middle) and leg damping ratio touchdown angle β (left), leg spring stiffness k ζ0 (right). Arrows indicate increasing directions for each varied parameter.

The dependence of equilibrium points on the leg stiffness, illustrated in the middle figure shows that increasing spring constants cause an increase in the range of stable energy levels. This is also natural since an increased stiffness corresponds to shorter stance times, resulting in decreased damping losses and a corresponding decrease in the necessary torque input. Finally, we observe that the impact of the damping coefficients on the equilibrium points is not as pronounced, providing evidence that our compensation strategy successfully balances damping losses. Nevertheless, increasing the amount of damping causes a slight decrease in the range of stable energy levels.

IV. F EEDBACK C ONTROL OF TD-SLIP RUNNING A. Deadbeat Control by Inversion of the Apex Return Map

D. Correspondence of the Model to Biological Data A recent quantitative comparison of ground reaction force data from a variety of running animals to those predicted by a simple, passive spring-mass model shows that despite the very good correspondence of vertical force components between biological data and the idealized SLIP model, there are some discrepancies in how well horizontal forces can be predicted [38]. In this section, we report on an interesting property of the torque-actuated TD-SLIP morphology: It seems to be capable of qualitatively reproducing ground reaction force profiles very similar to those observed in biological systems. Fig. 6 illustrates the body trajectory for a single stride of steady-state running with the TD-SLIP model, together with a depiction of “virtual footfalls” in the direction of instantaneous ground reaction force vectors throughout the stance phase. As a result of the ramp torque profile we use for supplying energy to the system, large backward horizontal forces are

The presence of a sufficiently accurate analytic formulation of the apex return map naturally motivates its inversion to obtain a controller for stabilizing the system around a desired operating point [ya∗ , x˙ ∗a ] in apex state coordinates. A similar approach was adopted in a number of studies [8, 29, 31], but never in the context of a lossy model or torque actuation. In this section, we describe a deadbeat gait controller for TD-SLIP as an application of our approximations, and show that it is capable of very accurately regulating the apex states of a running TD-SLIP and improves on both the accuracy and stability of previous attempts to control a similar, torqueactuated model in [12]. An explicitly specified desired apex state will require a nonzero change in the energy level of the system. Using a strategy similar to the energy-conserving torque controller of Section III-A, we will use the hip torque to supply the requested energy input to the system in a single step. Similar to (21), this energy is given by 1 m((x˙ ∗a )2 − x˙ 2a ) + mg(ya∗ − ya ) + Eloss , (22) 2 which can easily be solved to determine the ramp torque magnitude τ0 , assuming, once again, that tf = tlo .



46

Eτ

=

Once the desired torque profile is determined, the return map has only one remaining degree of control freedom: the touchdown angle θtd . A deadbeat controller can be formulated as a one dimensional minimization problem in the form θtd = argmin −π 2
@ 7\SLFDOO\ /L ZLOO EH D QRQVLQJXODU VHULDO FRPELQDWLRQ RI OL RQH 'R) MRLQWV 7KH À[DWLRQ FDQ EH DQ HPEHGPHQW OL =  RU FDQ UHOHDVH VHYHUDO 'R)V VXFK WKDW ∀L ∈ {, .., Q} ,  ≤ OL ≤  .



,QGHHG FKRRVLQJ OL ≥  ZRXOG FRUUHVSRQG WR FRPSOHWH IUHHGRP EHWZHHQ HL DQG RL ZKLFK ZRXOG QRW PDNH DQ\ SUDFWLFDO VHQVH LQ WKH FRQVLGHUHG DSSOLFDWLRQ ZKHUH IRUFH WUDQVPLVVLRQ LV UHTXLUHG %HWZHHQ RL− DQG RL  RQ WKH URERW VLGH WKHUH LV DQ DFWLYH PHFKDQLVP 5L ZKLFK FRQQHFWLYLW\ LV GHQRWHG UL  6LPLODUO\ EHWZHHQ HL− DQG HL RQ WKH KXPDQ VLGH WKHUH LV D PHFKDQLVP +L RI FRQQHFWLYLW\ KL  1RWH WKDW GXH WR WKH FRPSOH[LW\ RI KXPDQ NLQHPDWLF KL LV QRW DOZD\V H[DFWO\ NQRZQ DQG OLWHUDWXUH IURP ELRPHFKDQLFV SURYLGHV FRQWURYHUVLDO GDWD RQ WKLV SRLQW )RU H[DPSOH WKH HOERZ LV RIWHQ PRGHOHG DV D RQH 'R) MRLQW EXW LQ UHDOLW\ D UHVLGXDO VHFRQG 'R) FDQ EH REVHUYHG >@ 2XU JRDO LV WR GHVLJQ PHFKDQLVPV /L ZLWK L ∈ {, .., Q} LQ VXFK D ZD\ WKDW RQ RQH VLGH DOO WKH IRUFHV JHQHUDWHG E\ WKH H[RVNHOHWRQ RQ WKH KXPDQ OLPE DUH FRQWUROODEOH DQG RQ WKH RWKHU VLGH WKHUH LV QR SRVVLEOH PRWLRQ IRU WKH H[RVNHOHWRQ ZKHQ WKH KXPDQ OLPE LV VWLOO :H VKDOO WKXV FRQVLGHU LQ WKH

 114

∀L ∈  · · · Q,

6Q 6Q

7L = {}

DQG

D

:/L → = {} ,

E

ZKHUH LV WKH VSDFH RI WZLVWV GHVFULELQJ WKH YHORFLWLHV RI URERW ERG\ RL UHODWLYH WR R ZKHQ WKH ZKROH PHFKDQLVP 6Q LV FRQVLGHUHG DQG 6Q :/L →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ÀUVW RQH FDQ QRWLFH WKH UHFXUVLYH VWUXFWXUH RI WKH FRQVLG HUHG V\VWHP LI ZH QDPH 6L WKH VXEPHFKDQLVP FRQVWLWXWHG E\ WKH ERGLHV R WR RL  WKH FKDLQV 5 WR 5L DQG / WR /L  ZH FDQ UHSUHVHQW 6L UHFXUVLYHO\ IURP 6L−  DV LQ )LJ  ZKHUH PL−

)LJ 

5HFXUVLYH VWUXFWXUH 6L RI WKH V\VWHP

LV WKH FRQQHFWLYLW\ RI 6L−  ,Q WKLV FRQYHQWLRQ 6 UHSUHVHQWV D

]HUR 'R) PHFKDQLVP 8VLQJ WKLV UHFXUVLYH UHSUHVHQWDWLRQ RQH FDQ HVWDEOLVK WKH IROORZLQJ SURSRVLWLRQ 3URSRVLWLRQ  7KH FRQGLWLRQV  DUH HTXLYDOHQW WR  ∀L ∈  · · · Q, GLP(76L− + 75L + 7/L ) = 

∀L ∈  · · · Q, GLP(76L− ∩ 75L ) = 

DQG

D

DQG

E

GLP(76Q ) =  ,

F

6M 7

ZKHUH 76 M = M LV WKH VSDFH RI WZLVWV GHVFULELQJ WKH YHORFLWLHV RI R M UHODWLYH WR R  ZKHQ 6 M LV FRQVLGHUHG LVRODWHG IURP WKH UHVW RI WKH PHFKDQLVP WKHQ LW LV GLIIHUHQW IURP 6Q 7 M  75L LV WKH VSDFH RI WZLVWV SURGXFHG E\ 5L ² LH WKH VSDFH RI WZLVWV RI RL UHODWLYH WR RL−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À[DWLRQ PHFKDQLVPV /L  2I FRXUVH DV LW LV XVXDO LQ PHFKDQLVP GHVLJQ ZKHQ D SDUWLFXODU GHVLJQ LV ÀQDOO\ SURSRVHG LW ZLOO EH QHFHVVDU\ WR YHULI\ D SRVWHULRUL WKH VLQJXODULW\ DYRLGDQFH FRQGLWLRQ

0RUHRYHU VLQFH LI $ DQG % DUH WZR YHFWRU VXEVSDFHV RI ( DQG GLP($) + GLP(%) > GLP(() WKHQ $ ∩ % = {} (T E LPSRVHV WKDW L−

∀L ∈  · · · Q, PL− + UL ≤  RU 

∑ (O M + U M ) + UL ≤ .L



M=

)LQDOO\ WKH ODVW FRQGLWLRQ F OHDGV WR Q

PQ =  RU 

∑ (O M + U M ) = .Q



M=

1RWLFH WKDW  SURYLGHV WKH WRWDO QXPEHU RI 'R)V WR EH IUHHG IRU WKH PHFKDQLVP 6Q  ZKLOH  JLYHV WKH PLQLPDO YDOXH WR SUHYHQW IURP K\SHUVWDWLFLW\ LQ WKH VXEPHFKDQLVPV 6 M IRU O M DQG  SURYLGHV WKH PD[LPDO RQH WR SUHYHQW IURP LQWHUQDO PRELOLW\ LQ 6 M  7KDQNV WR WKHVH WKUHH ODVW QHFHVVDU\ FRQGLWLRQV ZH DUH DEOH WR FDOFXODWH WKH GLIIHUHQW SRVVLEOH VROXWLRQV IRU GLVWULEXWLQJ WKH DGGLWLRQDO SDVVLYH 'R)V RYHU WKH VWUXFWXUH • WKH SRVVLEOH FKRLFHV IRU O DUH VXFK WKDW  ≥ O ≥  − U  • IRU HDFK FKRLFH RI O  WKH SRVVLEOH FKRLFHV IRU O DUH VXFK WKDW  ≥ O ≥  − U − U − O  7KLV OHDGV WR D WUHH WKDW JURXSV DOO WKH DGPLVVLEOH FRPELQDWLRQV IRU OL  DV LOOXVWUDWHG LQ )LJ  

& &RQGLWLRQV RQ FRQQHFWLYLWLHV $W ÀUVW OHW·V FRPSXWH WKH FRQQHFWLYLW\ RI 6L  2QH KDV 76L = 7/L ∩ (75L + 76L− ) ,



ZKLFK GLUHFWO\ UHVXOWV IURP WKH VSDFH VXP ODZ IRU VHULDO FKDLQV DQG WKH LQWHUVHFWLRQ ODZ IRU SDUDOOHO FKDLQV VHH >@  )XU WKHUPRUH VLQFH IRU DQ\ YHFWRU VXEVSDFHV $ DQG % GLP($) + GLP(%) = GLP($ + %) + GLP($ ∩ %) RQH JHWV PL

= GLP(7/L ) + GLP(75L + 76L− ) − GLP(7/L + 75L + 76L− ) = GLP(7/L ) + GLP(75L ) + GLP(76L− ) − GLP(75L ∩ 76L− ) −GLP(7/L + 75L + 76L− ).

,I FRQGLWLRQ  LV UHVSHFWHG DQG XQGHU IXOO UDQN DVXPSWLRQ RQH JHWV  PL = OL + UL + PL− −  )LQDOO\ XVLQJ P =  WKLV UHFXUVLYH HTXDWLRQ VLPSOLÀHV WR L

PL =

∑ (O M + U M ) − .L

.



M=

7KHUHIRUH IURP (T D  QRWLFLQJ WKDW DQ\ YHFWRU VXEVSDFHV $% DQG & RI D YHFWRU VSDFH ( GLP($ + % + &) ≤ GLP($) + GLP(%) + GLP(&) LW LV QHFHVVDU\ WKDW L

∀L ∈  · · · Q, PL− + UL + OL ≥ ,

RU 

∑ (O M + U M ) ≥ .L



M=

 115

)LJ  7UHH RI SRVVLEOH VROXWLRQV IRU WKH QXPEHU RI SDVVLYH 'R)V WR DGG DW HYHU\ À[DWLRQ SRLQW

2XW RI WKLV WUHH DOO WKH SRVVLEOH FRPELQDWLRQV RI FRQQHFWLYLWLHV IRU WKH À[DWLRQV DUH JLYHQ 2I FRXUVH WKH VHOHFWLRQ DPRQJ WKHVH VROXWLRQV LV WR EH PDGH GHSHQGLQJ RQ WKH H[RVNHOHWRQ NLQHPDWLFV *HQHUDOO\ VSHDNLQJ DQ LPSRUWDQW DVSHFW WR EH FRQVLGHUHG LV WKH IRUFH WUDQVPLVVLRQ WKURXJK DQ\ OLQHDU RU URWDWLRQDO 'R) WKDW LV QRW IUHHG E\ WKH À[DWLRQ PHFKDQLVP D IRUFH RU D PRPHQW ZLOO EH WUDQVPLWWHG WR WKH KXPDQ OLPE ZKLFK LV VXUURXQGHG E\ VRIW WLVVXHV 7KHUHIRUH W\SLFDOO\ WUDQVPLWWLQJ PRPHQWV DURXQG 3L ZRXOG OHDG WR ORFDOO\ GHIRUP WKH WLVVXHV ZKLFK LQ WXUQ FDQ JHQHUDWH GLVFRPIRUW 7KH QH[W VHFWLRQ LOOXVWUDWHV RQ D FRQFUHWH VSDWLDO H[DPSOH LQYROYLQJ WZR À[DWLRQV KRZ WR LQWHJUDWH WKLV NLQG RI FRQVLGHUDWLRQV LQ WKH GHVLJQ RI À[DWLRQ PHFKDQLVPV ,,, $ 33/,&$7,21 72 $ *,9(1 (;26.(/(721 $ $%/( DQ XSSHU OLPE H[RVNHOHWRQ IRU UHKDELOLWDWLRQ $%/( VHH )LJ  LV D  D[LV H[RVNHOHWRQ WKDW KDV EHHQ GHVLJQHG E\ &($/,67 >@ RQ WKH EDVLV RI DQ LQQRYDWLYH VFUHZDQGFDEOH DFWXDWLRQ WHFKQRORJ\ >@  ,WV NLQHPDWLFV LV FRPSRVHG RI D VKRXOGHU VSKHULFDO MRLQW FRPSRVHG RI  FRLQFLGHQW SLYRWV DQG D  'R) SLYRW HOERZ 7KH IRUHDUP

#                   

             

$  %







     !"

% & & '

 

 

 

' (









)LJ 

.LQHPDWLFV RI $%/(

WHUPLQDWHG E\ D KDQGOH LV QRW DFWXDWHG 'HWDLOV RQ WKLV URERW FDQ EH IRXQG LQ >@ % )L[DWLRQV GHVLJQ IRU $%/( ,Q WKLV VHFWLRQ ZH DSSO\ JHQHUDO PHWKRG SURSRVHG LQ 6HF ,, WR $%/( :H SURFHHG LQ WKUHH VWHSV

ZD\ RI DFKLHYLQJ JOREDO LVRVWDWLFLW\ 'HJUHHV RI )UHHGRP IRU / KDYH WR EH FKRVHQ FRPSOHPHQWDU\ WR WKRVH RI 5 LQ RUGHU WR VDWLVI\ WKH IXOO UDQN DVVXPSWLRQ 6LQFH 5 LV D EDOO MRLQW WKDW JHQHUDWHV WKUHH LQGHSHQGHQW URWDWLRQDO YHORFLWLHV DURXQG LWV FHQWHU 0  / PXVW JHQHUDWH WKUHH LQGHSHQGHQW YHORFLWLHV DW SRLQW 0  )RU H[DPSOH WKUHH QRQ FRSODQDU WUDQVODWLRQV FRXOG EH XVHG IRU /  +RZHYHU LQ WKLV FDVH WKH À[DWLRQ ZRXOG WUDQVPLW D QXOO IRUFH LH D SXUH FRXSOH 7KLV VHHPV XQGHVLUDEOH GXH WR WKH WRUVLRQ RI WKH VRIW WLVVXHV WKDW LW ZRXOG FUHDWH DURXQG 3 DW WKH OHYHO RI WKH DWWDFKPHQW WR WKH OLPE 2QH FRXOG WKXV WKLQN RI XVLQJ IRU / D EDOO MRLQW DURXQG 3  EXW LQ WKLV FDVH WKH IXOO UDQN FRQGLWLRQ ZRXOG QRW EH UHVSHFWHG EHFDXVH 5 DQG / ZRXOG ERWK JHQHUDWH WKH VDPH −−−→ URWDWLRQ DURXQG ] = −−−→ 0 3  )LQDOO\ D SUHIHUUHG VROXWLRQ 0 3  LV WR FKRRVH IRU / WZR SLYRW MRLQWV SHUSHQGLFXODU WR WKH DUP PDLQ D[LV ]DUP  DQG RQH WUDQVODWLRQ MRLQW FROOLQHDU ]DUP VHH  IRU IXUWKHU GHÀQLWLRQ RI 3L  ]DUP DQG ] I RUHDUP  ,Q WKLV FDVH WZR IRUFHV SHUSHQGLFXODU WR]DUP DQG RQH PRPHQW DURXQG ]DUP FDQ EH H[FKDQJHG EHWZHHQ WKH H[RVNHOHWRQ DQG WKH DUP WKURXJK /  0RUHRYHU VLQFH 6 LV LVRVWDWLF RQH KDV P =  7KHUHIRUH / QHHGV WR EH GHVLJQHG LQ RUGHU WR EH NLQHPDWLFDOO\ FRPSOHPHQWDU\ WR 5  ZKLFK LV D SLYRW RI D[LV (0 ,] ) 6HH $SSHQGL[ IRU ] GHÀQLWLRQ  ,Q RWKHU ZRUGV / PXVW JHQHUDWH LQGHSHQGHQWO\  URWDWLRQV SHUSHQGLFXODU WR ] DQG  YHORFLWLHV DW SRLQW 0  $ QDWXUDO VROXWLRQ LV WR FKRRVH D EDOO MRLQW DURXQG 3 DQG WZR WUDQVODWLRQV LQ D SODQH SHUSHQGLFXODU WR ]  7KH UHVXOWLQJ RYHUDOO GHVLJQ LV QRWHG D DQG UHSUHVHQWHG LQ )LJ  

)LJ 

6FKHPDWLF RI WKH $%/( DQG KXPDQ DUP FRXSOLQJ

 

FRPSXWH WKH WUHH RI SRVVLEOH YDOXHV IRU OL • FKRRVH DPRQJ WKHP D SUHIHUUHG VROXWLRQ E\ H[DPLQLQJ IRUFH WUDQVPLVVLRQ SURSHUWLHV DQG NLQHPDWLF FRPSOHPHQ WDULW\ • YHULI\ WKH IXOO NLQHPDWLF UDQN ZKLFK LV UHSRUWHG LQ $S SHQGL[ % )LUVWO\ VLQFH $%/( FRPSULVHV DQ XSSHU DUP DQG D IRUHDUP ZH VKDOO XVH WZR À[DWLRQV 6HH )LJ   7KH WRWDO QXPEHU RI SDVVLYH 'R) WR EH DGGHG LV JLYHQ E\ (T   •

Q=

Q=

∑ O M =  − ∑ U M =  − ( + )

M=

M=

⇒ O + O = 



0RUHRYHU IRU WKH ÀUVW À[DWLRQ (T  DQG  JLYH  − U ≤ O ≤ 

⇒

 ≤ O ≤  .

6LQFH WKH WRWDO QXPEHU RI 'R)V LV À[HG WKH WUHH RI SRVVLEOH VROXWLRQV FRQVLVWV KHUH RI WKUHH SDUDOOHO EUDQFKHV ZKHUH O LV FKRVHQ EHWZHHQ  DQG  DQG O =  − O  3RVVLEOH FRXSOHV IRU (O , O ) DUH    DQG   +HUHDIWHU WKHVH WKUHH RSWLRQV DUH DQDO\]HG LQ RUGHU WR FKRRVH D SUHIHUUHG GHVLJQ DPRQJ WKHP • &DVH D O =  DQG O =  ,Q WKLV FDVH ERWK 6 WDNHQ LVRODWH DQG 6 DUH LVRVWDWLF ZKLFK FRUUHVSRQGV WR WKH PRVW LQWXLWLYH



116

 





 




 

 







 

 




      

  

 

 

)LJ  &RQVLGHUHG SRVVLELOLWLHV IRU FRXSOLQJ $%/( WR DQ KXPDQ DUP &DVH D  EDOO MRLQW DORQH DW 3 DQG EDOO MRLQW   VOLGHV DW 3  FDVH E  8QLYHUVDO MRLQW  VOLGH DW 3 DQG EDOO MRLQW   VOLGHV DW 3  FDVH F %DOO MRLQWV ZLWK VOLGH DW ERWK 3 DQG 3 

•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

VKRZQ LQ $SSHQGL[ % WKDW VLQJXODU FRQÀJXUDWLRQV RI WKLV VROX WLRQ QRWHG E DQG UHSUHVHQWHG LQ )LJ  DUH HDVLO\ LGHQWLÀDEOH DQG IDU DZD\ IURP QRPLQDO FRQGLWLRQV RI RSHUDWLRQ • &DVH F O =  DQG O =  6LPLODUO\ WR VROXWLRQ D  WKLV FRPELQDWLRQ ZLOO QHFHVVDU\ OHDG WR WUDQVPLW DW OHDVW RQH WRUVLRQ PRPHQW DURXQG ] I RUHDUP  DV LOOXVWUDWHG LQ )LJ  VROXWLRQ F  7KHUHIRUH WKH ÀQDOO\ SUHIHUUHG VROXWLRQ LV E  1RWH WKDW ZLWK VROXWLRQ E  JHQHUDWLQJ D PRPHQW WR WKH KXPDQ XSSHU DUP DURXQG ]DUP LV REWDLQHG E\ DSSO\LQJ RSSRVLWH SXUH IRUFHV SHUSHQGLFXODU WR ]DUP DW 3 DQG WR ] I RUHDUP DW 3 VHH

'R) 5RWDWLRQ ⊥ WR WKH OLPE D[LV 5RWDWLRQ ⊥ WR WKH OLPE D[LV 5RWDWLRQ DURXQG WKH OLPE D[LV 7UDQVODWLRQ

$UP )L[DWLRQ ◦ ◦ ◦ PP

)RUHDUP )L[DWLRQ ◦ ◦ ◦ PP

7$%/( ,

DW 3 DQG 3 UHVSHFWLYHO\ )RU WKH H[SHULPHQWV SUHVHQWHG LQ WKH QH[W VHFWLRQ LQ RUGHU WR FRPSDUH WKH IRUFHV LQYROYHG ZLWK DQG ZLWKRXW 'R) OLEHUDWLRQ WKH À[DWLRQV ZHUH DOVR HTXLSSHG ZLWK UHPRYDEOH PHWDOOLF SLQV DOORZLQJ XV WR TXLFNO\ ORFN WKH SDVVLYH 'R) ZLWKRXW GHWDFKLQJ WKH VXEMHFW IURP WKH H[RVNHOHWRQ 7KHVH À[DWLRQV

)LJ  7UDQVPLWWLQJ D PRPHQW DURXQG WKH XSSHU DUP D[LV ZLWK VROXWLRQ E OHIW DQG F ULJKW

)LJ   ,QWHUHVWLQJO\ WKLV UHSURGXFHV WKH PHWKRG XVHG E\ SK\VLFDO WKHUDSLVWV WR DVVLVW SDWLHQWV LQ JHQHUDWLQJ LQWHUQDO URWDWLRQV RI WKH VKRXOGHU ZLWKRXW WRUVLRQ WR WKH WLVVXH $V D SULFH WKH IXOO H[WHQVLRQ FRQÀJXUDWLRQ ZKHQ 0  3 DQG 3 DUH DOLJQHG LV VLQJXODU DV GHWDLOHG LQ WKH $SSHQGL[ % 7KLV FRQÀJXUDWLRQ FDQ EH HDVLO\ DYRLGHG E\ OLPLWLQJ WKH UDQJH RI WKH HOERZ H[WHQVLRQ & )L[DWLRQV UHDOL]DWLRQ 7KH WZR À[DWLRQV PHFKDQLVPV DUH ÀQDOO\ LGHQWLFDO 7KH\ VKDOO JHQHUDWH WKUHH LQGHSHQGHQW URWDWLRQV DQG RQH WUDQVODWLRQ DORQJ WKH OLPE 7KH PHFKDQLVP XVHG WR UHDOL]H WKLV IXQFWLRQ

)LJ 

7KH WZR À[DWLRQV RQ WKH H[RVNHOHWRQ

ZHUH PRXQWHG RQ WKH  'R) $%/( H[RVNHOHWRQ $UP À[DWLRQ LV SODFHG QHDU WKH HOERZ MXVW XQGHU WKH WULFHSV )RUHDUP À[DWLRQ LV SODFHG QHDU WKH ZULVW 7KHUPRIRUPDEOH PDWHULDOV ZHUH DOVR XVHG WR FUHDWH WZR VSOLQWV SHUIHFWO\ DGDSWHG WR KXPDQ PRUSKRORJ\ 7KHVH VSOLQWV DUH VHULDOO\ FRQQHFWHG WR WKH ODVW À[DWLRQ ERG\ :ULVW VSOLQW ZDV VSHFLÀFDOO\ FUHDWHG WR ORFN WKH ZULVW ÁH[LRQV ZKLFK DUH QRW VWXGLHG KHUH 2QO\ SDVVLYH SURQRVXSLQDWLRQ LV DOORZHG ,9 ( ;3(5,0(17$/ 5(68/76 $ ([SHULPHQWDO VHWXS

)LJ 

)L[DWLRQ VLPSOLÀFDWLRQ DQG UHDOL]DWLRQ UHDU DQG IURQW

FRQVLVWV RI WKUHH VXFFHVVLYH SLYRW MRLQWV ZKLFK D[LV FRLQFLGH DQG RQH VOLGHU ZKRVH D[LV LV SDUUDOOHO WR KXPDQ OLPE VHH )LJ   7KH À[DWLRQV ZHUH GLPHQVLRQQHG GLIIHUHQWO\ RQH WR DOORZ IRUHDUP SURQRVXSLQDWLRQ DQG WKH RWKHU QRW WR FROOLGH ZLWK DUP WLVVXHV $V D UHVXOW SRVVLEOH PRWLRQV OHIW E\ WKH SDVVLYH À[DWLRQV KDYH WKH UDQJHV GHÀQHG LQ 7DEOH , 7KHVH À[DWLRQV ZHUH ERWK ÀWWHG ZLWK RQH IRUFH VHQVRU SODFHG RQ WKH EDVH $7, 1DQR D[LV )RUFH7RUTXH VHQVRU DOORZLQJ XV WR UHFRQVWUXFW WKH WKUHH IRUFHV DQG WKUHH WRUTXHV FRPSRQHQWV

 117

$Q DUWLFXODWHG PDQQHTXLQ ZDV XVHG IRU WKH H[SHULPHQW ,WV DUPV SRVVHVV  SDVVLYH 'R)V D EDOO MRLQW VKRXOGHU D SLYRW HOERZ DQG D SURQRVXSLQDWLRQ  $QDO\]LQJ WKH LQWHUDFWLRQ IRUFH DQG WRUTXH YDULDWLRQV DW WKH LQWHUIDFHV GXULQJ WKH VDPH PRYHPHQW ZLWK LVRVWDWLF À[DWLRQV DQG ZLWKRXW ORFNHG FDVH ZLOO DOORZ XV WR HYDOXDWH WKHLU LPSDFW RQ SUHYHQWLQJ WKH DSSHDUDQFH RI XQFRQWUROOHG IRUFHV EXW DOVR WR TXDQWLI\ WKHP URXJKO\ 7KH PDQQHTXLQ ZDV WKXV SODFHG LQ WKH H[RVNHOHWRQ DQG DW WDFKHG WKURXJK WKH IZR À[DWLRQV 7KH WKHUPRIRUPDEOH VSOLQWV DOORZ WR DYRLG DQ\ ORRVHQHVV LQ WKH À[DWLRQ DQG LQFUHDVH WKH FRQWDFW VWLIIQHVV QR IRDP QHHGHG  'XULQJ WKH H[SHULPHQWV WKH H[RVNHOHWRQ LPSRVHV D FRQWUROOHG WUDMHFWRU\ ZLWK D FRQVWDQW VSHHG WR WKH PDQQHTXLQ DUP 7KH H[SHULPHQW FRQVLVWV LQ D VHULH RI VL[ VLPSOH SRLQWWRSRLQW PRYHPHQWV ZLWK D OLPLWHG UDQJH RI PRWLRQ WR WKH VDPH WDUJHW EXW UHDFKHG ZLWK GLIIHUHQW MRLQWV PRYHPHQWV WKDQNV

          

 










       


       

  )LJ  $OORZHG IRUFHV  )[ + )\ QRUP RQ WKH WZR À[DWLRQ PHDQ IRU WKH VL[ PRYHPHQWV )LJ 

0DQQHTXLQ FRQQHFWHG WR WKH $%/( H[RVNHOHWRQ

WR DUP UHGXQGDQF\  7DUJHW ZDV UHDFKHG DW FRQVWDQW DQG ORZ VSHHG  PV WR OLPLW LQHUWLDO IRUFHV 'XH WR WKH ULJLGLW\ RI WKH PDQQHTXLQ VXUIDFH WKH PRYHPHQW DPSOLWXGH RQ HYHU\ H[RVNHOHWRQ MRLQW ZDV OLPLWHG WR D  ◦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À[DWLRQ  :H FDQ REVHUYH RQ WKH DUP 

     

 












 

 

  

   

)LJ  YDOXH RI WKH XQGHVLUHG IRUFH |)[ | DQG PRPHQWV  $YHUDJHG DEVROXWH  0[ + 0\ + 0] RQ WKH WZR À[DWLRQV PHDQ IRU WKH VL[ PRYHPHQWV QRUP

À[DWLRQ D GHFUHDVH LQ WKH XQGHVLUHG IRUFH )[ OHYHO DQG WKH XQGHVLUHG WRUTXHV OHYHO E\ DSSUR[LPDWLYHO\  )RU WKH IRUHDUP À[DWLRQ DQ DSSUR[LPDWLYH  GHFUHDVH FDQ EH REVHUYHG IRU WKH XQGHVLUHG IRUFH DQG PRPHQW FRPSRQHQWV ,Q )LJ  WKH QRUP RI WKH FRPSRQHQWV )\ DQG )] FRUUHVSRQGLQJ WR WKH FRPSRQHQWV WUDQVPLWWHG E\ WKH SDVVLYH À[DWLRQV LV SUHVHQWHG 7KH H[RVNHOHWRQ DELOLW\ WR WUDQVPLW IRUFHV WR WKH VXEMHFW LV QRW DOWHUHG 1RWH WKDW WKH SHUFHQWDJH RI K\SHUVWDWLF IRUFH OHYHO GHFUHDVH DFKLHYHG E\ WKH À[DWLRQV UHVXOWLQJ IURP RXU PHWKRGRORJ\ DQG WKH REWDLQHG QXPHULFDO YDOXH RI WKH

 118

K\SHUVWDWLF IRUFHV KDYH WR EH LQWHUSUHWHG ,QGHHG GXH WR WKH PDQQHTXLQ DUP VPDOOQHVV VHH )LJ  DQG LW ERG\ VXIDFH ULJLGLW\ K\SHUVWDWLF IRUFH OHYHO LV KLJKHU WKDQ WKH RQH RFFXULQJ GXULQJ D FRPDQLSXODWLRQ EHWZHHQ WKH H[RVNHOHWRQ DQG D KXPDQ VXEMHFW 9 & 21&/86,21 ,Q WKLV SDSHU ZH SUHVHQWHG D PHWKRGRORJ\ DLPHG DW GHVLJQ LQJ WKH NLQHPDWLFV RI À[DWLRQV EHWZHHQ DQ H[RVNHOHWRQ DQG D KXPDQ PHPEHU 7KDQNV WR WKLV PHWKRG ZH UHDOL]HG LVRVWDWLF À[DWLRQV IRU D  'R) H[RVNHOHWRQ DQG H[SHULPHQWDOO\ YHULÀHG WKHLU EHQHÀW RQ PLQLPL]LQJ XQFRQWUROODEOH K\SHUVWDWLF IRUFHV DW WKH KXPDQ URERW LQWHUIDFH DQG WKXV RQ D ÀQH FRQWURO RI WKH LQWHUDFWLRQ IRUFHV 7KHVH UHVXOWV VKRZ WKDW WKH SURYLGHG VROXWLRQ DYRLGV K\SHUVWDWLFLW\ EXW DOVR DGDSWV WR ODUJH YDULDWLRQV RI WKH KXPDQ OLPE JHRPHWU\ ZLWKRXW UHTXLULQJ D FRPSOH[ DGDSWDEOH URERW VWUXFWXUH $ 33(1',; $ 'HPRQVWUDWLRQ RI 3URSRVLWLRQ   &RQGLWLRQV  DUH VXIÀFLHQW [() ⇒ ()] :H KHUH VXSSRVH WKDW FRQGLWLRQV  DUH YHULÀHG %HFDXVH LQ 6Q  RL− LV FRQQHFWHG RQ RQH VLGH WR R WKURXJK 6L− DQG RQ WKH RWKHU VLGH WR RL WKURXJK 5L VHH )LJ   RQH KDV / 0 ∀L ∈ { . . . Q}, 6Q 7L− =6L− 7L− ∩ 75L + 6Q 7L , 

ZKLFK LV D UHFXUVLYH UHODWLRQVKLS IRU 6Q 7L  5HFDOOLQJ WKDW E\ DVVXPSWLRQ 6Q 76Q = {} FRQGLWLRQ F DQG 76L− ∩ 75L = {} FRQGLWLRQ E  WKLV UHFXUVLYH ODZ WULYLDOO\ OHDGV WR D  )XUWKHUPRUH WKH NLQHPDWRVWDWLF GXDOLW\ SULQFLSOH DSSOLHG WR WKH ORRS (R → RL− → RL → R ) LQ )LJ  ZULWHV

∀L ∈ { . . . Q}, GLP(6L :/L → ) + GLP(76L− + 75L + 7/L ) =  .  7KDQNV WR FRQGLWLRQ D  WKLV OHDGV WR ∀L ∈ { . . . Q},

6L

:/L → = {} .



&RQVLGHULQJ DJDLQ WKH V\VWHP 6L GHSLFWHG LQ )LJ  DQG UHFDOO LQJ WKDW /L DQG 5L DUH VHULDO FKDLQV RQH KDV ∀L ∈ { . . . Q} 6L

:/L → =6L :/L →L =6L :5L →L =6L :5L →L− = {} .



7KHUHIRUH VWDWLFDOO\ VSHDNLQJ WKH PXOWLORRS V\VWHP 6L− LV LQ WKH VDPH VWDWH ZKHQ LQFOXGHG LQ 6L WKDQ ZKHQ LVRODWHG IURP WKH UHVW RI WKH PHFKDQLVP ∀L ∈ { . . . Q},

6L

:/L− → =6L− :/L− → ,

ZKLFK WRJHWKHU ZLWK  UHFXUVLYHO\ / OHDGV WR 0FRQGLWLRQ E   &RQGLWLRQV  DUH QHFHVVDU\  () ⇒ () 

)LUVWO\ LI FRQGLWLRQ F LV QRW YHULÀHG WKHQ 6Q 7Q = 76Q = {} ,Q WKLV FDVH D LV QRW VDWLVÀHG 6HFRQGO\ LI E LV QRW YHULÀHG WKHQ ∃L, (75L ∩ 76L− ) = {} 7KDQNV WR (T   WKLV OHDGV WR ∃L ∈ { · · · Q},

6Q

7L− = {} ,



W′ =

% 6LQJXODULW\ DQDO\VLV IRU $%/( DQG WKH WZR SURSRVHG À[DWLRQ PHFKDQLVPV :H VWXG\ WKH PHFKDQLVP GHSLFWHG LQ )LJ  5 LV D EDOO MRLQW ZKLFK FHQWHU LV 0  / LV FRPSRVHG RI D EDOO MRLQW ZKLFK −−−→ − FHQWHU LV 3 ZLWK 0 3 = O .→ ] DQG O =  DQG D VOLGH DORQJ → − − → (3 , ]DUP ) 5 LV D SLYRW MRLQW ZKLFK D[LV LV (0 , [ ) / LV −−−→ − FRPSRVHG RI D EDOO MRLQW ZKLFK FHQWHU LV 3 ZLWK 0 3 = O .→ ] − − − − → DQG O =  DQG D VOLGH DORQJ (3 , ] I RUHDUP ) ,Q RUGHU WR ÀQG WKH VLQJXODU FRQÀJXUDWLRQV RI WKLV V\VWHP ZH XVH WKH QHFHVVDU\ DQG VXIÀFLHQW FRQGLWLRQV  

  O

6LQFH O =  τ = {W , ..,W } LV D EDVLV RI 2 1 LI DQG RQO\ LI τ = W , ..,W ,W′ ,W′ LV D EDVLV RI R  /HW·V FRQVLGHU QRZ DL ∈ R, L ∈ {, .., } VXFK WKDW ZLWK GHW($) =

DW + DW + DW + DW + DW′ + DW′ = 



WKHQ ∃L, 6L :/L→ = {} PHDQLQJ WKDW 6L WDNHQ LVRODWH LV K\SHUVWDWLF 2EYLRXVO\ DGGLQJ WKH UHVW RI WKH PHFKDQLVP WR EXLOG 6Q  ZKLFK FRQVLVWV RI DGGLQJ D SDUDOOHO EUDQFK WR 6L EHWZHHQ R DQG RL ZLOO QRW GHFUHDVH WKH GHJUHH RI K\SHUVWDWLFLW\ 7KHUHIRUH ∃L, 6Q :/L→ = {} ZKLFK FRQWUDGLFWV FRQGLWLRQ E 

(W − W ) (W − W ) = ( 7 [ 7 )7 DQG W′ = = ( 7 \ 7 )7 , O O

ZH FDQ HDVLO\ VKRZ WKDW   W W W W W′ W′ = $ [W W W W W W ]

R

ZKLFK GLUHFWO\ FRQWUDGLFWV D  7KLUGO\ LI D LV QRW YHULÀHG LH ∃L, GLP(76L− + 75L + 7/L ) ≤  ,

7KXV 75 + 7/ = VSDQ {W , ...,W } 'HÀQLQJ



,W LV WULYLDO WR VKRZ WKDW D = D = D =  D G] =  D + − → → − → − → D G\ =  DQG D + D G[ =  ZKHUH ]−− DUP = G[ [ + G\ \ + G] ]  ,I G] =  WKHQ D = D = D =  7KHUHIRUH τ DQG τ DUH EDVHV RI R  2WKHUZLVH WKHUH H[LVWV D QRQ QXOO FRPELQDWLRQ RI DL WKDW YHULÀHV  ZKLFK PHDQV WKDW τ DQG τ DUH QRW IUHH DQ\PRUH &RQGLWLRQ D LV WKXV YHULÀHG IRU L =  LI DQG RQO\ − →→ π LI ]−− DUP . ] =  7KLV LV HTXLYDOHQW WR α = ±  DQG WKLV LV D VLQJXODU YDOXH WR EH DYRLGHG ,Q WKH UHVW RI WKH VWXG\ ZH ZLOO − →→ WKXV FRQVLGHU WKDW ]−− DUP . ] =  • )RU L =  D ZULWHV GLP(76 + 75 + 7/ ) =  :H NQRZ WKDW 76 = 75 ∩ 7/  /HW·V FRQVLGHU W ∈ 7/ DQG W ′ ∈ 75  2QH KDV 

∃(α , α , α , α )

VXFK WKDW W = ∑ αL WL

∃(α′ , α′ , α′ , ) VXFK WKDW



L= W ′ = α′ W + α′ W + α′ W

→→ − 8VLQJ ]−− DUP . ] =  RQH HDVLO\ JHWV W = W ′ ⇔ α = α = α = α′ = α′ =  .



W = W ′ ⇔ W = α W = α′ W .



RU

,Q RWKHU ZRUGV DW SRLQW 3  76 = 75 ∩ 7/ = VSDQ({W }) = VSDQ({(] 7  7 )7 }) .  :H NQRZ ZULWH WZLVWV DW SRLQW 3  :H JHW 76 = VSDQ({W }) 75 = VSDQ({W }) DQG 7/ = VSDQ({W W W W }) ZLWK )LJ  .LQHPDWLFV RI $%/(  LWV À[DWLRQV 7KH SODQH RI WKH ÀJXUH SHUSHQGLFXODU WR [  LV GHÀQHG E\ 0  3 DQG 3 ZKLOH 0 LV RII WKH SODQH

 ([DPLQDWLRQ RI &RQGLWLRQ D •

)RU L =  D ZULWHV GLP(75 + 7/ ) = 

$W SRLQW 3  YHORFLWLHV DOORZHG E\ / EHORQJ WR WKH YHFWRU VXEVSDFH 7/ = VSDQ {W ,W ,W ,W } DQG WKH YHORFLWLHV DOORZHG E\ 5 EHORQJ WR 75 = VSDQ {W ,W ,W } ZLWK W = ([ 7  7 )7 , W = (] 7  7 )7 , W = ([ 7 − O .\ 7 )7

W = (\ 7  7 )7 , W = ( 7 ]DUP 7 )7 , W = (\ 7 O .[ 7 )7

 119

W = (] 7 O VLQ θ [ 7 )7 , W = ([ 7 − O \ 7 )7 , W = ([ 7 7 )7

W = (\ 7 7 )7 , W = (] 7 7 )7 , W = (7 ] I RUHDUP 7 )7 ,   −−→ − −] PHDVXUHG DURXQG [  7KXV  ] , → ZKHUH 3 3 = O] DQG θ = →  76 + 75 + 7/ = VSDQ({W ,W ,W ,W ,W ,W }) → − 6XSSRVH ÀUVW WKDW VLQ θ =  7KHQ GHQRWLQJ − ] = ][ .→ [ + − − \ + ]] .→ ]  RQH JHWV ]\ .→ W = ][W + ]\W + ]]W



,Q WKLV SDUWLFXODU FDVH {W .. W } LV QRW D EDVLV ZKLFK LGHQWLÀHV D VHFRQG VLQJXODU FRQÀJXUDWLRQ ZKHQ 0  3 DQG 3 DUH DOLJQHG ,Q WKH UHVW RI WKH VWXG\ ZH ZLOO WKXV DVVXPH WKDW WKLV

VLQJXODU FRQÀJXUDWLRQ LV DOVR DYRLGHG WKDW LV VLQ θ =  'HÀQLQJ (W − ][W − ]\W − ]]W ) = (7 [ 7 )7 , DQG O VLQ θ (W − W ) = (7 \ 7 )7 , W′ = O   − ′ = %. [W W .. W ] ZLWK GHW(%) = O VLQ ZH JHW W′ W′ W′ .. W θ =    7KXV τ = {W .. W } LV D EDVLV RI R LI DQG RQO\ LI τ = ′ } LV D EDVLV RI R  /HWV·V FRQVLGHU E ∈ R, L ∈ {, .., } {W′ .. W L VXFK WKDW W′

=

EW′ + EW′ + EW + EW + EW + EW =  .



′ EW′ + EW′ + EW

= ,W FRPHV HDVLO\ WKDW E = E = E =  DQG − → − −−→ →  ZKLFK LV HTXLYDOHQW WR E − [ + E → \ + E ]−− I RUHDUP =   7KH QHFHVVDU\ DQG VXIÀFLHQW FRQGLWLRQV WR KDYH D QRQQXOO WULSOHW − − −−→ [ , → \ , ]−− E , E , E YHULI\LQJ WKH SUHYLRXV HTXDWLRQ LV WKDW → I RUHDUP DUH FRSODQDU 7KLV LGHQWLÀHV D WKLUG VLQJXODULW\ ZKLFK DJDLQ LV VXSSRVHG WR EH DYRLGHG LQ WKH UHVW RI WKH VWXG\  ([DPLQDWLRQ RI WKH FRQGLWLRQ E • )RU L =  VLQFH 76 = {} RQH GLUHFWO\ JHWV GLP(76 ∩ 7/ ) =  • )RU L =  LW LV QHFHVVDU\ WR YHULI\ WKDW GLP(76 ∩ 7/ ) =  /HW·V FRQVLGHU W ∈ 76 DQG W ′ ∈ 7/  2QH KDV ∃α ∈ R ∃α′ , α′ , α′ , α′ ∈ R

/ W = αW / W ′ = α′ W + α′ W + α′ W + α′ W .

2QH HDVLO\ VKRZV WKDW W = W ′ LV HTXLYDOHQW WR 3 − −−→ → − α O VLQ θ → [ + α′ ]−− I RUHDUP =  → − − − − (α ][ + α′ )→ [ + (α ]\ + α′ )→ \ + (α ]] + α′ )→ ] =  − −−→ 6LQFH → [ LV QRW FROLQHDU WR ]−− I RUHDUP  WKH ÀUVW HTXDWLRQ OHDGV − − − ′ [ ,→ \ ,→ ] } IRUPV D EDVLV WR α = α =  6LPLODUO\ VLQFH {→ 









α′ = α′ = α′ =  ,Q FRQFOXVLRQ GLP(76 ∩ 7/ ) = {}  ([DPLQDWLRQ RI WKH FRQGLWLRQ F )RU WKH FRQVLGHUHG H[DPSOH Q =  DQG FRQGLWLRQ F ZULWHV GLP(76 ) =  6LQFH 76 = (76 + 75 ) ∩ 7/  ZH QHHG WR YHULI\ WKDW DQ\ YHFWRU WKDW EHORQJV WR ERWK (76 + 75 ) DQG 7/ LV QXOO /HW·V FRQVLGHU W ∈ (76 + 75 ) DQG W ′ ∈ 7/  2QH KDV ∃ α , α ∈ R

∃ α′ , .., α′ ∈ R

/ W = αW + αW / W ′ = α′ W + α′ W  + α′ W  + α′ W 

7KHUHIRUH W = W ′ LV HTXLYDOHQW WR 3 → − − −−→ − [ − α O → \ + α′ ]−− α O VLQ θ → I RUHDUP =  → − − − − (α ][ + α′ + α )→ [ + (α ]\ + α′ )→ \ + (α ]] + α′ )→ ] =  7KH ÀUVW RI WKHVH WZR HTXDWLRQV OHDGV WR α = α = α′ =  − − −−→ VLQFH LW LV VXSSRVHG WKDW → [ → \ DQG ]−− I RUHDUP DUH QRW FRSODQDU LQ RUGHU WR DYRLG WKH WKLUG VLQJXODULW\ DQG VLQ θ =  LQ RUGHU WR DYRLG WKH VHFRQG VLQJXODULW\ 7KHUHIRUH WKH VHFRQG HTXDWLRQ − − − [ , → \ , → ] } IRUPV D EDVLV OHDGV WR α = α = α′ =  EHFDXVH {→ ′ ,Q FRQFOXVLRQ W = W ⇒ W =  RU GLP(76 ) =   6XPPDU\ ,Q FRQFOXVLRQ ZH LGHQWLÀHG WKUHH VLQJXODULWLHV

 120

→→ −  − ]− DUP . ] =  UHSUHVHQWLQJ WKH FDVH ZKHUH WKH SDVVLYH VOLGH PRXQWHG SDUDOOHO WR WKH XSSHU DUP D[LV LV SHUSHQGLFXODU WR WKH URERW XSSHU OLPE D[LV 7KLV FDVH ZLOO QHYHU → − → DSSHDU VLQFH WKH DQJOH EHWZHHQ ]−− DUP DQG ] UHÁHFWV VPDOO GLVFUHSDQFLHV EHWZHHQ WKH H[RVNHOHWRQ DQG KXPDQ NLQHPDWLFV DQG UHPDLQV VPDOOHU WKDQ D IHZ GHJUHHV  VLQ θ =  UHSUHVHQWLQJ WKH FDVH ZKHUH 0  3 DQG 3 DUH DOLJQHG 7KLV VLQJXODU FRQÀJXUDWLRQ FDQ EH DYRLGHG E\ OLPLWLQJ WKH UDQJH RI PRWLRQ IRU WKH URERW HOERZ WR D IHZ GHJUHHV EHIRUH IXOO H[WHQVLRQ − − −−→  → [ → \ DQG ]−− I RUHDUP FRSODQDU 7KLV FRQÀJXUDWLRQ GRHV QRW DSSHDU LQ SUDFWLFH VLQFH LQ WKH QRPLQDO FRQÀJXUDWLRQ → → − \ DQG [ LV SHUSHQGLFXODU WR WKH SODQH JHQHUDWHG E\ − − − − − → ] I RUHDUP  7KHUHIRUH XQGHU QRUPDO FRQGLWLRQV RI RSHUDWLRQ WKH $%/( H[RVNHOHWRQ ZLWK LWV WZR À[DWLRQV QHYHU IDOOV LQWR D VLQJXODU FRQÀJXUDWLRQ 5 ()(5(1&(6 >@ $% =RVV + .D]HURRQL DQG $ &KX %LRPHFKDQLFDO GHVLJQ RI WKH EHUNHOH\ ORZHU H[WUHPLW\ H[RVNHOHWRQ EOHH[  0HFKDWURQLFV ,((($60( 7UDQVDFWLRQV RQ   ² DSULO  >@ 0 0LKHOM 7 1HI DQG 5 5LHQHU $UPLQ LL   GRI UHKDELOLWDWLRQ URERW PHFKDQLFV DQG NLQHPDWLFV ,Q 5RERWLFV DQG $XWRPDWLRQ  ,((( ,QWHUQDWLRQDO &RQIHUHQFH RQ SDJHV  ²   >@ -& 3HUU\ - 5RVHQ DQG 6 %XUQV 8SSHUOLPE SRZHUHG H[RVNHOHWRQ GHVLJQ 0HFKDWURQLFV ,((($60( 7UDQVDFWLRQV RQ   ² DXJ  >@ -RV / 3RQV :HDUDEOH 5RERWV %LRPHFKDWURQLF ([RVNHOHWRQV :LOH\ $SULO  >@ 6WHSKHQ + 6FRWW DQG 'DYLG $ :LQWHU %LRPHFKDQLFDO PRGHO RI WKH KXPDQ IRRW .LQHPDWLFV DQG NLQHWLFV GXULQJ WKH VWDQFH SKDVH RI ZDONLQJ -RXUQDO RI %LRPHFKDQLFV  ² 6HSWHPEHU  >@ )&7 9DQ GHU +HOP +(- 9HHJHU *0 3URQN /+9 9DQ GHU :RXGH DQG 5+ 5R]HQGDO *HRPHWU\ SDUDPHWHUV IRU PXVFXORVNHOHWDO PRGHOOLQJ RI WKH VKRXOGHU V\VWHP -RXUQDO RI %LRPHFKDQLFV  ²  )HEUXDU\  >@ $ 6FKLHOH $Q H[SOLFLW PRGHO WR SUHGLFW DQG LQWHUSUHW FRQVWUDLQW IRUFH FUHDWLRQ LQ SKUL ZLWK H[RVNHOHWRQV ,Q 5RERWLFV DQG $XWRPDWLRQ  ,&5$  ,((( ,QWHUQDWLRQDO &RQIHUHQFH RQ SDJHV  ²    >@ $ 6FKLHOH DQG )&7 YDQ GHU +HOP .LQHPDWLF GHVLJQ WR LPSURYH HUJRQRPLFV LQ KXPDQ PDFKLQH LQWHUDFWLRQ 1HXUDO 6\VWHPV DQG 5H KDELOLWDWLRQ (QJLQHHULQJ ,((( 7UDQVDFWLRQV RQ   ² GHF  >@ / : /DPRUHX[ .LQHPDWLF PHDVXUHPHQWV LQ WKH VWXG\ RI KXPDQ ZDONLQJ %XOO 3URVWKHW 5HV  ²  30,'  >@ ./ 0DUNROI -6 0HQVFK DQG +& $PVWXW] 6WLIIQHVV DQG OD[LW\ RI WKH NQHH²WKH FRQWULEXWLRQV RI WKH VXSSRUWLQJ VWUXFWXUHV D TXDQWLWDWLYH LQ YLWUR VWXG\ - %RQH -RLQW 6XUJ $P  ²  >@ & 'LH]0DUWQH] - 5LFR - &HUYDQWHV6QFKH] DQG - *DOODUGR 0R ELOLW\ DQG FRQQHFWLYLW\ LQ PXOWLORRS OLQNDJHV ,Q $GYDQFHV LQ 5RERW .LQHPDWLFV SDJHV ²  >@ 0 6WRNGLMN & * 0 0HVNHUV + ( - 9HHJHU @ >@ DQG OHDYH LVVXHV RI XQFHUWDLQW\ WR WKH FRQWURO SKDVH LQ ZKLFK WKH SDWK PD\ EH H[HFXWHG XVLQJ D IHHGEDFN FRQWUROOHU >@ 3ODQQLQJ DQG FRQWURO DUH UHODWHG EXW GLVWLQFW ÀHOGV :KLOH UHFHQW ZRUN RQ SDWK SODQQLQJ KDV DGGUHVVHG PRWLRQ DQGRU VHQVLQJ XQFHUWDLQW\ VHH 6HFWLRQ ,,  PRVW SODQQLQJ PHWKRGV GR QRW DFFRXQW IRU FRQWURO GXULQJ H[HFXWLRQ DQG PRVW FRQWURO PHWKRGV WDNH WKH SDWK DV JLYHQ /4*03 EXLOGV D EULGJH EHWZHHQ WKHVH GLVFLSOLQHV DQG GUDZV IURP UHVXOWV LQ ERWK 7KLV ZRUN ZDV VXSSRUWHG LQ SDUW E\ 16) $ZDUG  DQG 1,+ $ZDUG 5(%$ 7KH DXWKRUV DUH ZLWK WKH 8QLYHUVLW\ RI &DOLIRUQLD DW %HUNHOH\ %HUNHOH\ &$ 86$ (PDLO {EHUJ SDEEHHO JROGEHUJ}#EHUNHOH\HGX

 129

.HQ *ROGEHUJ

ct

D

E

)LJ  D 7KH PD[LPXP IDFWRU ct E\ ZKLFK WKH HOOLSVH FRQWDLQLQJ WKH SRVLWLRQV ZLWKLQ RQH VWDQGDUG GHYLDWLRQ FDQ EH VFDOHG EHIRUH LW LQWHUVHFWV REVWDFOHV JLYHV DQ LQGLFDWLRQ RI WKH SUREDELOLW\ WKDW FROOLVLRQV ZLOO EH DYRLGHG WRS  ct LV FRPSXWHG DV WKH (XFOLGHDQ GLVWDQFH WR WKH QHDUHVW REVWDFOH LQ WKH HQYLURQPHQW WUDQVIRUPHG VXFK WKDW WKH HOOLSVH EHFRPHV D XQLW GLVF ERWWRP  E 7KH HOOLSVHV VKRZ WKH DSULRUL GLVWULEXWLRQV DV FRPSXWHG E\ /4*03 DORQJ WKH EHVW DPRQJ WKH  FDQGLGDWH SDWKV IRU 6FHQDULR $ 7KH VDPSOHV UHVXOW IURP SHUIRUPLQJ  VLPXODWLRQV

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À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ÀFDQW LQÁXHQFH RQ ZKLFK SDWK LV EHVW $ SDWK SODQQHU WKDW LV XQDZDUH RI WKH VHQVRUV WKH FRQWUROOHU DQG WKHLU XQFHUWDLQWLHV ZRXOG QRW EH DEOH WR PDNH WKLV GLVWLQFWLRQ DQG PD\ SURGXFH VXERSWLPDO SDWKV 7KH UHPDLQGHU RI WKLV SDSHU LV RUJDQL]HG DV IROORZV :H VWDUW E\ GLVFXVVLQJ UHODWHG ZRUN LQ 6HFWLRQ ,, :H IRUPDOO\ GHÀ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ÀUVW WZR EXW LV WR VRPH H[WHQW DOVR DSSOLFDEOH WR WKH ODWWHU DV ZH ZLOO VKRZ LQ RQH RI RXU H[SHULPHQWV 3ODQQHUV WKDW VSHFLÀFDOO\ WDNH LQWR DFFRXQW PRWLRQ XQFHU WDLQW\ LQFOXGH >@ >@ >@ 7KHVH SODQQHUV SODQ SDWKV WKDW DYRLG URXJK WHUUDLQ EXW GR QRW FRQVLGHU SDUWLDO REVHUYDELOLW\ DQG VHQVLQJ XQFHUWDLQW\ ,Q >@ WKH SUREDELOLW\ RI FROOLVLRQV LV PLQLPL]HG IRU WKH VSHFLÀF FDVH RI D PDQLSXODWRU ZLWK EDVH SRVH XQFHUWDLQW\ 7KH VHQVLQJ XQFHUWDLQW\ LV WDNHQ LQWR DFFRXQW LQ WKH SODQQHU RI >@ ZKLFK DLPV WR RSWLPL]H WKH LQIRUPDWLRQ FRQWHQW DORQJ D SDWK 3ODQQHUV LQ >@ >@ >@ DVVXPH WKDW ODQGPDUN UHJLRQV H[LVW LQ WKH HQYLURQPHQW ZKHUH WKH DFFXPXODWHG PRWLRQ XQFHUWDLQW\ FDQ EH ´UHVHWµ 2WKHU DSSURDFKHV EOHQG SODQQLQJ DQG FRQWURO E\ GHÀ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

 130

PLQLPL]H H[SHFWHG FRVW >@ >@ >@ >@ +RZHYHU WKHVH DOJRULWKPV LPSOLFLWO\ DVVXPH WR UHFHLYH PD[LPXPOLNHOLKRRG PHDVXUHPHQWV IURP WKH VHQVRUV ZKLFK GRHV QRW UHVXOW LQ WKH WUXH SUREDELOLW\ GLVWULEXWLRQV RI WKH VWDWH RI WKH URERW EXW UDWKHU D PHDVXUH RI KRZ ZHOO RQH ZLOO EH DEOH WR LQIHU WKH VWDWH %HVLGHV WKH VHQVRUV /4*03 DOVR WDNHV LQWR DFFRXQW WKH FRQWUROOHU WKDW ZLOO EH XVHG IRU H[HFXWLQJ WKH SDWK DQG FRPSXWHV WKH WUXH DSULRUL SUREDELOLW\ GLVWULEXWLRQV RI WKH VWDWH RI WKH URERW DORQJ LWV IXWXUH SDWK ,,, 3 52%/(0 ' (),1,7,21 /HW X = Rn EH WKH VWDWH VSDFH RI WKH URERW DQG OHW U = Rm EH WKH FRQWURO LQSXW VSDFH RI WKH URERW :H DVVXPH WKDW WLPH LV GLVFUHWL]HG LQWR VWDJHV RI HTXDO GXUDWLRQ DQG WKDW DSSO\LQJ D FRQWURO LQSXW ut ∈ U DW VWDJH t EULQJV WKH URERW IURP VWDWH xt ∈ X DW VWDJH t WR VWDWH xt+1 ∈ X DW VWDJH t + 1 DFFRUGLQJ WR D JLYHQ VWRFKDVWLF G\QDPLFV PRGHO xt = f (xt−1 , ut−1 , mt ),

mt ∼ N (0, Mt ),



ZKHUH mt LV WKH SURFHVV QRLVH DW VWDJH t GUDZQ IURP D ]HUR PHDQ *DXVVLDQ GLVWULEXWLRQ ZLWK YDULDQFH Mt WKDW PRGHOV WKH PRWLRQ XQFHUWDLQW\ :H DVVXPH WKDW WKH IXQFWLRQ f LV HLWKHU OLQHDU RU ORFDOO\ ZHOO DSSUR[LPDWHG E\ LWV OLQHDUL]DWLRQ /HW XV EH JLYHQ D VWDUW VWDWH xstart ∈ X ZKHUH WKH URERW EHJLQV DQG D JRDO UHJLRQ X goal ⊂ X ZKHUH WKH URERW QHHGV WR JR $ SDWK Π IRU WKH URERW LV GHÀQHG DV D VHULHV RI VWDWHV DQG FRQWURO LQSXWV (x⋆0 , u⋆0 , . . . , x⋆ℓ , u⋆ℓ ) VXFK WKDW x⋆0 = xstart  x⋆ℓ ∈ X goal  DQG x⋆t = f (x⋆t−1 , u⋆t−1 , 0) IRU 0 < t ≤ ℓ ZKHUH ℓ LV WKH QXPEHU RI VWDJHV RI WKH SDWK 7KDW LV D SDWK FRQQHFWV WKH VWDUW VWDWH DQG WKH JRDO UHJLRQ DQG LV FRQVLVWHQW ZLWK WKH G\QDPLFV PRGHO LI WKHUH ZHUH QR SURFHVV QRLVH 'XULQJ H[HFXWLRQ RI WKH SDWK WKH URERW ZLOO GHYLDWH IURP WKH SDWK GXH WR PRWLRQ XQFHUWDLQW\ 7R FRPSHQVDWH IRU XQH[SHFWHG PRWLRQV ZH DVVXPH WKDW WKH SDWK ZLOO EH H[HFXWHG XVLQJ D IHHGEDFN FRQWUROOHU WKDW DLPV WR NHHS WKH URERW FORVH WR WKH SDWK E\ PLQLPL]LQJ WKH FRVW IXQFWLRQ ℓ   E ((xt −x⋆t )TC(xt −x⋆t )+(ut −u⋆t )TD(ut −u⋆t )) ,  t=0

ZKLFK TXDGUDWLFDOO\ SHQDOL]HV GHYLDWLRQV IURP WKH SDWK C DQG D DUH JLYHQ SRVLWLYHGHÀQLWH ZHLJKW PDWULFHV :H DVVXPH WKDW QRLV\ VHQVRUV SURYLGH XV ZLWK SDUWLDO LQIRUPDWLRQ DERXW WKH VWDWH GXULQJ H[HFXWLRQ RI WKH SDWK DFFRUGLQJ WR D JLYHQ VWRFKDVWLF REVHUYDWLRQ PRGHO zt = h(xt , nt ),

nt ∼ N (0, Nt ),



ZKHUH zt LV WKH PHDVXUHPHQW REWDLQHG DW VWDJH t WKDW UHODWHV WR VWDWH xt WKURXJK IXQFWLRQ h DQG nt LV WKH PHDVXUHPHQW QRLVH GUDZQ IURP D ]HURPHDQ *DXVVLDQ ZLWK YDULDQFH Nt  :H DVVXPH WKDW WKH IXQFWLRQ h LV HLWKHU OLQHDU RU ORFDOO\ ZHOO DSSUR[LPDWHG E\ LWV OLQHDUL]DWLRQ :H GHÀQH RXU SUREOHP LQ WZR SDUWV L JLYHQ WKH VWRFKDVWLF G\QDPLFV PRGHO WKH VWRFKDVWLF REVHUYDWLRQ PRGHO DQG WKH FRVW IXQFWLRQ FRPSXWH WKH DSULRUL GLVWULEXWLRQV RI WKH VWDWH DQG FRQWURO LQSXW DORQJ D JLYHQ SDWK DQG LL JLYHQ D SODQQLQJ REMHFWLYH EDVHG RQ WKH SUREDELOLW\ GLVWULEXWLRQV VHOHFW WKH EHVW SDWK DPRQJ D ODUJH VHW RI FDQGLGDWHV

,9 $ 35,25, 3 52%$%,/,7< ' ,675,%87,216 ,Q WKLV VHFWLRQ ZH GHVFULEH KRZ WR FRPSXWH WKH DSULRUL SUREDELOLW\ GLVWULEXWLRQV RI WKH VWDWH DQG FRQWURO LQSXW RI WKH URERW DORQJ D JLYHQ SDWK Π )RU WKLV ZH XVH WKH IDFW WKDW ZH NQRZ LQ DGYDQFH ZKDW FRQWUROOHU ZLOO EH XVHG WR H[HFXWH WKH SDWK IRU OLQHDU G\QDPLFV DQG REVHUYDWLRQ PRGHOV ZLWK *DXVVLDQ QRLVH DQG D TXDGUDWLF FRVW IXQFWLRQ WKH RSWLPDO DSSURDFK IRU H[HFXWLQJ WKH SDWK LV WR XVH DQ /45 IHHGEDFN FRQWUROOHU LQ SDUDOOHO ZLWK D .DOPDQ ÀOWHU IRU VWDWH HVWLPDWLRQ ZKLFK LV FDOOHG OLQHDUTXDGUDWLF *DXVVLDQ /4* FRQWURO >@ $ .DOPDQ ÀOWHU SURYLGHV WKH RSWLPDO HVWLPDWH RI WKH VWDWH JLYHQ SUHYLRXV VWDWH HVWLPDWHV PHDVXUHPHQWV DQG FRQWURO LQSXWV DQG DQ /45 FRQWUROOHU SURYLGHV WKH RSWLPDO FRQWURO LQSXW JLYHQ WKH HVWLPDWH RI WKH VWDWH :H ZLOO ÀUVW GLVFXVV KRZ WR OLQHDUL]H WKH G\QDPLFV DQG REVHUYDWLRQ PRGHO DQG WKHQ UHYLHZ WKH .DOPDQ ÀOWHU DQG /45 FRQWUROOHU )URP WKHVH ZH FRPSXWH WKH DSULRUL SURED ELOLW\ GLVWULEXWLRQV RI WKH VWDWHV DQG WKH FRQWURO LQSXWV RI WKH URERW DORQJ WKH SDWK $ /LQHDUL]HG '\QDPLFV DQG 2EVHUYDWLRQ 0RGHO ,Q SULQFLSOH RXU DSSURDFK DSSOLHV WR OLQHDU G\QDPLFV DQG REVHUYDWLRQ PRGHOV f DQG h +RZHYHU VLQFH WKH URERW LV FRQWUROOHG WR VWD\ FORVH WR WKH SDWK GXULQJ H[HFXWLRQ ZH FDQ DSSUR[LPDWH QRQOLQHDU PRGHOV ZLWK ORFDO OLQHDUL]DWLRQV LH ÀUVWRUGHU 7D\ORU H[SDQVLRQV DURXQG WKH SDWK Π 7KLV JLYHV WKH IROORZLQJ OLQHDUL]HG VWRFKDVWLF PRGHOV xt = f (x⋆t−1 , u⋆t−1 , 0) + At (xt−1 − x⋆t−1 ) + Bt (ut−1 − u⋆t−1 ) + Vt mt , zt = h(x⋆t , 0) + Ht (xt − x⋆t ) + Wt nt ,

 

∂f ∂f ZKHUH At = ∂x (x⋆t−1 , u⋆t−1 , 0) Bt = ∂u (x⋆t−1 , u⋆t−1 , 0) ∂f ∂h ∂h ⋆ ⋆ ⋆ (x⋆t , 0) Vt = ∂m (xt−1 , ut−1 , 0) Ht = ∂x (xt , 0) Wt = ∂n DUH WKH -DFRELDQ PDWULFHV RI f DQG h DORQJ SDWK Π ,W LV FRQYHQLHQW WR H[SUHVV WKH FRQWURO SUREOHP LQ WHUPV RI WKH GHYLDWLRQ IURP WKH SDWK %\ GHÀQLQJ

¯ t = xt − x

x⋆t ,

¯ t = ut − u

u⋆t ,

¯zt = zt −

h(x⋆t , 0),



DV WKH VWDWH GHYLDWLRQ FRQWURO LQSXW GHYLDWLRQ DQG PHDVXUH PHQW GHYLDWLRQ UHVSHFWLYHO\ ZH FDQ IRUPXODWH WKH G\QDPLFV DQG REVHUYDWLRQ PRGHO RI (TXDWLRQV  DQG  DV ¯ t−1 + Bt u ¯ t−1 + Vt mt , ¯ t = At x x ¯ ¯ zt = Ht xt + Wt nt ,

mt ∼ N (0, Mt ), nt ∼ N (0, Nt ),

 

DQG WKH FRVW IXQFWLRQ RI (TXDWLRQ  DV ℓ   ¯t + u ¯ TtD¯ (¯ xTt C x ut ) . E



t=0

7KLV LV WKH VWDQGDUG IRUPXODWLRQ RI DQ /4*FRQWURO SUREOHP % .DOPDQ )LOWHU IRU 2SWLPDO 6WDWH (VWLPDWLRQ ˜ t DQG 7KH .DOPDQ ÀOWHU NHHSV WUDFN RI WKH HVWLPDWH x ¯ t GXULQJ WKH H[HFXWLRQ RI WKH YDULDQFH Pt RI WKH WUXH VWDWH x SDWK ,W FRQWLQXDOO\ SHUIRUPV WZR VWHSV D SURFHVV XSGDWH WR

 131

¯ t  DQG D PHDVXUHPHQW SURSDJDWH WKH DSSOLHG FRQWURO LQSXW u XSGDWH WR LQFRUSRUDWH WKH REWDLQHG PHDVXUHPHQW z¯t  3URFHVV XSGDWH ˜− ˜ t−1 + Bt u ¯ t−1 x t = At x



Pt− = At Pt−1 ATt + Vt Mt VtT ,



0HDVXUHPHQW XSGDWH Kt = Pt− HtT (Ht Pt− HtT + Wt Nt WtT )−1



˜t = x ˜− ˜− x zt − H t x t + Kt (¯ t )



Pt = (I − Kt Ht )Pt− .



7KHVH DUH WKH VWDQGDUG .DOPDQ ÀOWHU HTXDWLRQV IRU RSWLPDO HVWLPDWLRQ JLYHQ WKH G\QDPLFV DQG REVHUYDWLRQ PRGHO RI (TXDWLRQV  DQG  >@ 1RWH WKDW WKH .DOPDQJDLQ PDWUL FHV Kt FDQ EH FRPSXWHG LQ DGYDQFH LH EHIRUH H[HFXWLRQ JLYHQ WKH LQLWLDO YDULDQFH P0  ZLWKRXW NQRZOHGJH RI WKH DFWXDO ¯ t DQG PHDVXUHPHQWV ¯ zt  FRQWURO LQSXWV u & /45 IRU 2SWLPDO &RQWURO ¯ t WKDW DUH RSWLPDO WR DSSO\ GXULQJ 7KH FRQWURO LQSXWV u H[HFXWLRQ RI WKH SDWK DUH GHWHUPLQHG E\ WKH FRQWURO SROLF\ WKDW PLQLPL]HV WKH FRVW IXQFWLRQ RI (TXDWLRQ   )RU WKH G\QDPLFV PRGHO RI (TXDWLRQ   WKH FRVW IXQFWLRQ LV PLQLPDO ¯ t  ZKHUH Lt LV WKH IHHGEDFN PDWUL[ ZKLFK LV ¯ t = Lt x ZKHQ u FRPSXWHG LQ DGYDQFH IRU DOO t ∈ 0, . . . , ℓ − 1 XVLQJ Sℓ = C Lt = St =



T T −(Bt+1 St+1 Bt+1 + D)−1 Bt+1 St+1 At+1 T T C + At+1 St+1 At+1 + At+1 St+1 Bt+1 Lt .

 

7KHVH DUH WKH VWDQGDUG HTXDWLRQV IRU D ÀQLWHKRUL]RQ GLVFUHWH WLPH /45 FRQWUROOHU >@ ˜ t RI WKH VWDWH ¯ t LV XQNQRZQ WKH HVWLPDWH x $V WKH WUXH VWDWH x ZKLFK LV REWDLQHG IURP WKH .DOPDQ ÀOWHU LV XVHG WR GHWHUPLQH ¯ t DW HDFK VWDJH t GXULQJ H[HFXWLRQ RI WKH WKH FRQWURO LQSXW u SDWK +HQFH WKH FRQWURO SROLF\ LV ¯ t = Lt x ˜t. u



$IWHU DSSOLFDWLRQ RI WKH FRQWURO LQSXW WKH .DOPDQ ÀOWHU SURGXFHV WKH HVWLPDWH RI WKH QH[W VWDWH IURP ZKLFK LQ WXUQ D QHZ FRQWURO LQSXW LV GHWHUPLQHG 7KLV F\FOH UHSHDWV XQWLO WKH H[HFXWLRQ RI WKH SDWK LV FRPSOHWH ' $SULRUL 'LVWULEXWLRQV RI 6WDWH DQG &RQWURO ,QSXW *LYHQ WKH /45 FRQWURO SROLF\ DQG WKH .DOPDQ ÀOWHU ¯ t DQG WKH ZH FDQ DQDO\]H LQ DGYDQFH KRZ WKH WUXH VWDWH x ˜ t ZLOO HYROYH GXULQJ H[HFXWLRQ RI WKH SDWK HVWLPDWHG VWDWH x DV IXQFWLRQV RI HDFK RWKHU 7KH HYROXWLRQ RI WKH WUXH VWDWH ¯ t LV GHSHQGHQW RQ WKH HVWLPDWHG VWDWH WKURXJK WKH /45 x FRQWURO SROLF\ (TXDWLRQ  DQG WKH HYROXWLRQ RI WKH ˜ t LV GHSHQGHQW RQ WKH WUXH VWDWH WKURXJK WKH HVWLPDWHG VWDWH x PHDVXUHPHQW REWDLQHG LQ WKH .DOPDQ ÀOWHU (TXDWLRQ   7KLV JLYHV WKH IROORZLQJ HTXDWLRQV ¯ t = At x ¯ t−1 + Bt Lt−1 x ˜ t−1 + Vt mt , x  ˜ t−1 + Bt Lt−1 x ˜ t−1 + Kt z¯t − ˜ t = At x x  ˜ t−1 + Bt Lt−1 x ˜ t−1 ) Ht (At x

 

 ˜ t−1 + Bt Lt−1 x ˜ t−1 + Kt Ht x ¯ t + Wt nt − = At x  ˜ ˜ Ht (At xt−1 + Bt Lt−1 xt−1 )  ˜ t−1 + Bt Lt−1 x ˜ t−1 + Kt = At x ¯ t−1 + Bt Lt−1 x ˜ t−1 + Vt mt ) + Wt nt − Ht (At x  ˜ t−1 + Bt Lt−1 x ˜ t−1 ) Ht (At x ˜ t−1 + Bt Lt−1 x ˜ t−1 + Kt Ht At x ¯ t−1 + = At x ˜ t−1 , Kt Ht Vt mt + Kt Wt nt − Kt Ht At x

X goal d φ v y

(TXDWLRQ  IROORZV IURP VXEVWLWXWLQJ (TXDWLRQ  LQWR (TXDWLRQ   7KH ÀUVW HTXDOLW\ RI  IROORZV IURP VXEVWL WXWLQJ (TXDWLRQ  LQWR (TXDWLRQ  DQG (TXDWLRQ  LQWR (TXDWLRQ   WKH VHFRQG DQG WKLUG HTXDOLWLHV IROORZ DIWHU VXEVWLWXWLQJ (TXDWLRQV  DQG   UHVSHFWLYHO\ DQG WKH IRXUWK HTXDOLW\ IROORZV DIWHU H[SDQGLQJ WKH WHUPV &RPELQLQJ (TXDWLRQV  DQG  JLYHV WKH PDWUL[ IRUP



¯ t−1 ¯t At x x Bt Lt−1 = + ˜t ˜ t−1 Kt Ht At At + Bt Lt−1 − Kt Ht At x x



Vt mt mt Mt 0 0 , ∼ N (0, ), nt Kt Ht Vt Kt Wt nt 0 Nt ZKLFK ZH ZULWH VKRUWKDQG IRU WKH DSSURSULDWH GHÀQLWLRQV RI yt  qt  Ft  Gt DQG Qt DV yt = Ft yt−1 + Gt qt ,

qt ∼ N (0, Qt ).



ˆ t DQG WKH YDULDQFH Rt )URP WKLV  ZH  FDQ FRPSXWH WKH PHDQ y RI yt = xx¯˜ tt IRU DQ\ VWDJH t RI WKH H[HFXWLRQ RI WKH SDWK ˆ t−1 , ˆ t = Ft y y

ˆ 0 = 0, y P R0 = 0 0



0 .  0   ˆ t LV ]HUR IRU DOO VWDJHV t +HQFH xx¯˜ tt ∼ 1RWH WKDW WKH PHDQ y N (0, Rt ) $V LW IROORZV IURP (TXDWLRQV  DQG  WKDW

⋆ ¯t xt I 0 x x = + t⋆ ,  ˜t ut ut 0 Lt x Rt = Ft Rt−1 FtT + Gt Qt GTt ,

WKH DSULRUL GLVWULEXWLRQ RI WKH VWDWH xt DQG WKH FRQWURO LQSXW ut DW VWDJH t RI WKH H[HFXWLRQ RI WKH SDWK LV

⋆ xt I 0 x Λt = .  ∼ N ( t⋆ , Λt Rt ΛTt ), ut ut 0 Lt x    7KH FRYDULDQFH EHWZHHQ uxii DQG ujj LV JLYHQ E\ x x T T cov( i , j ) = Λi Ri Fi+1 Fi+2 · · · FjT ΛTj , i < j.  ui uj

8VLQJ WKH DSULRUL GLVWULEXWLRQV WKH TXDOLW\ RI SDWK Π FDQ EH FRPSXWHG ZLWK UHVSHFW WR WKH FKRVHQ SODQQLQJ REMHFWLYH :H FDQ WKHQ XVH DQ\ PRWLRQ SODQQHU WR JHQHUDWH D ODUJH VHW RI FDQGLGDWH SDWKV IURP ZKLFK WKH EHVW RQH LV VHOHFWHG 9 ( ;$03/( $ 33/,&$7,216

$1'

5 (68/76

,Q WKLV VHFWLRQ ZH UHSRUW VLPXODWLRQ UHVXOWV IRU WKUHH VFHQDULRV LQ ZKLFK /4*03 LV XVHG WR VHOHFW D SDWK ,Q HDFK RI WKH WKUHH VFHQDULRV ZH XVH D GLIIHUHQW G\QDPLFV PRGHO REVHUYDWLRQ PRGHO DQG SODQQLQJ REMHFWLYH DQG SUR YLGH FRPSDUDWLYH DQDO\VLV ZLWK D EUXWHIRUFH DSSURDFK :H UHSRUW UHVXOWV IRU DQ ,QWHO 3 *+] ZLWK *% 5$0

 132

x

(x, y)

xstart D

θ

E

)LJ  D 7KH HQYLURQPHQW RI 6FHQDULR $ LQ ZKLFK D FDUOLNH URERW KDV WR PRYH EHWZHHQ D VWDUW VWDWH DQG D JRDO UHJLRQ ZLWKRXW FROOLGLQJ ZLWK REVWDFOHV 6HQVRUV FDQ RQO\ PHDVXUH WKH yFRRUGLQDWH RI WKH SRVLWLRQ RI WKH URERW 7KH EHVW SDWK DFFRUGLQJ WR /4*03 DPRQJ WKH  JHQHUDWHG E\ 557 LV VKRZQ E 7KH VWDWH x RI D FDUOLNH URERW

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ÀUVW VFHQDULR ZH DSSO\ /4*03 WR D QRQKROR QRPLF FDUOLNH URERW ZLWK QGRUGHU G\QDPLFV LQ D ' HQYLURQPHQW ZLWK REVWDFOHV 7KH URERW QHHGV WR PRYH IURP D VWDUW VWDWH xstart WR D JRDO UHJLRQ X goal ZLWKRXW FROOLGLQJ ZLWK WKH REVWDFOHV LQ WKH HQYLURQPHQW VHH )LJ D   '\QDPLFV PRGHO 7KH VWDWH x = (x, y, θ, v) RI WKH URERW LV D ' YHFWRU FRQVLVWLQJ RI LWV SRVLWLRQ (x, y) LWV RULHQWDWLRQ θ DQG LWV VSHHG v VHH )LJ E  ,WV FRQWURO LQSXW u = (a, φ) LV D ' YHFWRU FRQVLVWLQJ RI DQ DFFHOHUDWLRQ a DQG WKH VWHHULQJ ZKHHO DQJOH φ FRUUXSWHG E\ SURFHVV QRLVH  2  ˜ ∼ N (0, σa 02 ) 7KLV JLYHV WKH IROORZLQJ QRQ m = (˜ a, φ) 0 σφ OLQHDU G\QDPLFV PRGHO ⎤ ⎡ x + τ v cos θ ⎥ ⎢ y + τ v sin θ ⎥  f (x, u, m) = ⎢ ˜ ⎦, ⎣θ + τ v tan(φ + φ)/d v + τ (a + a ˜)

ZKHUH τ LV WKH GXUDWLRQ RI D VWDJH WLPH VWHS  DQG d WKH GLVWDQFH EHWZHHQ WKH IURQW DQG UHDU D[OH RI WKH FDU >@  2EVHUYDWLRQ PRGHO 7R VKRZ WKH HIIHFW RI SDUWLDO VHQVLQJ WKH URERW RQO\ UHFHLYHV IHHGEDFN RQ WKH yFRRUGLQDWH RI LWV SRVLWLRQ +HQFH WKH PHDVXUHPHQW YHFWRU z LV XQLYDULDWH DQG FRQVLVWV RI D PHDVXUHPHQW RI WKH yFRRUGLQDWH RI WKH URERW FRUUXSWHG E\ PHDVXUHPHQW QRLVH n = y˜ ∼ N (0, σy2 ) 7KLV JLYHV WKH IROORZLQJ OLQHDU REVHUYDWLRQ PRGHO h(x, n) = y + y˜.



(YHQ WKRXJK WKH VHQVRU IHHGEDFN LV YHU\ SDUWLDO LQIRUPD WLRQ DERXW WKH RWKHU YDULDEOHV LV VWLOO REWDLQHG WKURXJK WKH LQWHUSOD\ ZLWK WKH G\QDPLFV PRGHO

 3ODQQLQJ REMHFWLYH :H DLP WR ÀQG WKH SDWK IRU WKH URERW ZLWK D PLQLPDO SUREDELOLW\ RI FROOLGLQJ ZLWK REVWDFOHV ,QVWHDG RI FRPSXWLQJ WKLV SUREDELOLW\ H[DFWO\ ZH ZLOO XVH DQ DSSUR[LPDWLRQ WKDW FDQ EH FRPSXWHG HIÀFLHQWO\ JLYHQ WKH SUREDELOLW\ GLVWULEXWLRQV DORQJ WKH SDWK 7R WKLV HQG ZH ORRN DW WKH QXPEHU RI VWDQGDUG GHYLDWLRQV WKDW RQH FDQ GHYLDWH IURP WKH SDWK EHIRUH WKH URERW PD\ FROOLGH ZLWK DQ REVWDFOH /HW WKLV QXPEHU EH GHQRWHG ct IRU VWDJH t DORQJ WKH SDWK )RU D PXOWLYDULDWH *DXVVLDQ GLVWULEXWLRQ RI GLPHQVLRQ n WKH SUREDELOLW\ WKDW D VDPSOH LV ZLWKLQ ct VWDQGDUG GHYLDWLRQV LV JLYHQ E\ Γ(n/2, c2t /2) ZKHUH Γ LV WKH UHJXODUL]HG *DPPD IXQFWLRQ >@ ,W SURYLGHV D ORZHU ERXQG RI WKH SUREDELOLW\ RI DYRLGLQJ FROOLVLRQV DW VWDJH t :H QRZ GHÀQH WKH TXDOLW\ RI D SDWK Π DV ℓ 6

Γ(n/2, c2t /2),

b1







b2

 b3

vl r

b4

 b5



d θ

(x, y) vr



 D

E

)LJ  D 7KH HQYLURQPHQW RI 6FHQDULR % LQ ZKLFK HLJKW URERWV KDYH WR PRYH WR WKHLU DQWLSRGDO SRVLWLRQ LQ WKH HQYLURQPHQW ZLWKRXW PXWXDO FROOL VLRQV 7KH QXPEHUV LQGLFDWH WKH SULRULW\ UDQN DVVLJQHG WR HDFK URERW )LYH EHDFRQV b1 , . . . , b5 VHQG RXW D VLJQDO ZKRVH VWUHQJWK GHFD\V TXDGUDWLFDOO\ ZLWK GLVWDQFH E 7KH VWDWH x RI WKH GLIIHUHQWLDOGULYH URERW



t=0

ZKLFK LV LQGLFDWLYH RI WKH SUREDELOLW\ WKDW FROOLVLRQV ZLOO EH DYRLGHG GXULQJ H[HFXWLRQ ,W LV WKH SODQQLQJ REMHFWLYH WR ÀQG D SDWK IRU ZKLFK WKLV PHDVXUH LV PD[LPDO 7KH YDOXH RI ct IRU VWDJH t LV FRPSXWHG DV IROORZV )RU VLPSOLFLW\ ZH DSSUR[LPDWH WKH JHRPHWU\ RI WKH FDU E\ D ERXQGLQJ GLVF VXFK WKDW LWV RULHQWDWLRQ KDV QR LQÁXHQFH RQ ZKHWKHU RU QRW WKH FDU LV FROOLGLQJ $OVR LWV VSHHG GRHV QRW LQÁXHQFH LWV FROOLVLRQ VWDWXV +HQFH ct LV GHWHUPLQHG E\ WKH GLVWULEXWLRQ N (pt , Σt ) RI WKH SRVLWLRQ RI WKH FDU LH n = 2  ZKLFK  x⋆LV WKH PDUJLQDO GLVWULEXWLRQ RI WKH ÀUVW WZR YDULDEOHV RI N ( ut⋆ , Λt Rt ΛTt ) DV FRPSXWHG LQ (TXDWLRQ   /HW Ut EH t D PDWUL[ VXFK WKDW Ut UtT = Σt  7KH VHW RI SRVLWLRQV ZLWKLQ RQH VWDQGDUG GHYLDWLRQ LV WKHQ DQ HOOLSVH FHQWHUHG DW WKH PHDQ pt REWDLQHG E\ WUDQVIRUPLQJ D XQLW GLVF E\ Ut  DQG ct LV WKH PD[LPXP IDFWRU E\ ZKLFK WKH HOOLSVH FDQ EH VFDOHG VXFK WKDW LW GRHV QRW LQWHUVHFW ZLWK REVWDFOHV VHH )LJ D  &RPSXWLQJ ct FDQ HIÀFLHQWO\ EH LPSOHPHQWHG XVLQJ D FROOLVLRQFKHFNHU WKDW LV FDSDEOH RI SHUIRUPLQJ GLVWDQFH FDO FXODWLRQV DQG OLQHDU WUDQVIRUPDWLRQV RQ WKH JHRPHWU\ IRU LQ VWDQFH 62/,' >@ 7UDQVIRUPLQJ WKH HQYLURQPHQW LQFOXGLQJ WKH URERW E\ Ut−1 VXFK WKDW WKH XQFHUWDLQW\ HOOLSVH EHFRPHV D XQLW GLVF VHH )LJ D  DQG FDOFXODWLQJ WKH (XFOLGHDQ GLVWDQFH EHWZHHQ WKH URERW DQG WKH QHDUHVW REVWDFOH LQ WKH WUDQVIRUPHG HQYLURQPHQW JLYHV WKH YDOXH RI ct IRU VWDJH t  5HVXOWV :H UDQGRPO\ JHQHUDWHG  SDWKV XVLQJ WKH 557 DOJRULWKP ZKLFK WRRN  VHFRQGV )RU HDFK RI WKH SDWKV ZH FRPSXWHG WKH DSULRUL SUREDELOLW\ GLVWULEXWLRQV DQG WKH PHDVXUH RI (TXDWLRQ   ZKLFK WRRN LQ WRWDO  VHFRQGV 7KH EHVW SDWK DPRQJ WKH  LV VKRZQ LQ )LJ D  ,W FDQ EH VHHQ WKDW WKH ´ORZHUULJKWµ SDVVDJH LV FKRVHQ WR JHW WR WKH JRDO 7KLV FDQ EH H[SODLQHG DV WKH XQFHUWDLQW\ ZLOO PDLQO\ EH LQ WKH xFRRUGLQDWH JLYHQ WKDW WKH VHQVRUV RQO\ SURYLGH IHHGEDFN RQ WKH yFRRUGLQDWH 7KH JHRPHWU\ RI WKH ORZHUULJKW SDVVDJH DOORZV IRU PRUH GHYLDWLRQ LQ WKH xGLUHFWLRQ WKDQ WKH XSSHUOHIW SDVVDJH ,QGHHG FKDQJLQJ WKH REVHUYDWLRQ PRGHO VXFK WKDW RQO\ WKH xFRRUGLQDWH LV PHDVXUHG UHVXOWV LQ D SDWK WKDW WDNHV WKH XSSHUOHIW SDVVDJH 7R YDOLGDWH RXU UHVXOWV ZH XVHG D EUXWHIRUFH DSSURDFK WR HVWLPDWH IRU HDFK SDWK WKH ´JURXQGWUXWKµ SUREDELOLW\

 133

WKDW LW ZLOO EH H[HFXWHG ZLWKRXW FROOLVLRQV :H SHUIRUPHG  VLPXODWLRQV RI H[HFXWLRQV RI WKH SDWK XVLQJ WKH /45 FRQWUROOHU DQG DQ H[WHQGHG .DOPDQ )LOWHU ZLWK DUWLÀ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ÀFDQW LQÁXHQFH RQ ZKLFK SDWK LV RSWLPDO HYHQ DV WKH HQYLURQPHQW LV V\PPHWULF ,Q )LJ E WKH VDPSOHV RI  VLPXODWLRQV DUH VKRZQ IRU WKH EHVW DPRQJ WKH  SDWKV DORQJ ZLWK WKH XQFHUWDLQW\ HOOLSVHV RI WKH DSULRUL SUREDELOLW\ GLVWULEXWLRQV DV FRPSXWHG E\ /4*03 $V FDQ EH VHHQ WKH VDPSOHV LQGHHG IROORZ WKH DSULRUL GLVWULEXWLRQV FRPSXWHG E\ /4*03 7KLV VKRZV WKDW DQ\ HUURU LQWURGXFHG LQWR /4*03 E\ WKH OLQHDUL]DWLRQ RI WKH G\QDPLFV PRGHO LV LQVLJQLÀ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x = (x, y, θ) RI HDFK URERW LV D ' YHFWRU FRQVLVWLQJ RI LWV SRVLWLRQ (x, y) DQG LWV

RULHQWDWLRQ θ VHH )LJ E  ,WV FRQWURO LQSXW u = (vl , vr ) LV D ' YHFWRU FRQVLVWLQJ RI WKH VSHHGV RI WKH OHIW DQG ULJKW ZKHHO UHVSHFWLYHO\ FRUUXSWHG E\ SURFHVV QRLVH m = (˜ vl , v˜r ) ∼ N (0, σv2 I) 7KLV JLYHV WKH IROORZLQJ QRQOLQHDU G\QDPLFV PRGHO ⎡ ⎤ x + 12 τ (vl + v˜l + vr + v˜r ) cos θ 1 f (x, u, m) = ⎣ y + 2 τ (vl + v˜l + vr + v˜r ) sin θ ⎦ ,  θ + τ (vr + v˜r − vl − v˜l )/d ZKHUH τ LV WKH WLPH VWHS DQG d WKH GLVWDQFH EHWZHHQ WKH OHIW DQG ULJKW ZKHHO RI WKH URERW >@  2EVHUYDWLRQ PRGHO 7KH URERWV UHFHLYH IHHGEDFN RQ WKHLU VWDWH IURP ÀYH EHDFRQV b1 , . . . , b5 VFDWWHUHG DURXQG WKH HQYLURQPHQW WKDW HDFK VHQG RXW DQ LGHQWLÀDEOH VLJQDO RI XQLW VWUHQJWK WKDW GHFD\V TXDGUDWLFDOO\ ZLWK WKH GLVWDQFH WR WKH EHDFRQ (DFK EHDFRQ bi KDV D NQRZQ ORFDWLRQ (ˇ xi , yˇi , 1) +HQFH WKH PHDVXUHPHQW YHFWRU z FRQVLVWV RI ÀYH UHDGLQJV RI VLJQDO VWUHQJWKV RQH IURP HDFK EHDFRQ FRUUXSWHG E\ PHDVXUHPHQW QRLVH n = (˜b1 , . . . , ˜b5 ) ∼ N (0, σb2 I) 7KLV JLYHV WKH IROORZLQJ QRQOLQHDU REVHUYDWLRQ PRGHO ⎡ ⎤ 1/((x − xˇ1 )2 + (y − yˇ1 )2 + 1) + ˜b1 ⎢ ⎥  h(x, n) = ⎣ ⎦ .   2 2 1/((x − xˇ5 ) + (y − yˇ5 ) + 1) + ˜b5

 3ODQQLQJ REMHFWLYH )RU HDFK URERW ZH DLP WR PLQL PL]H WKH SUREDELOLW\ WKDW LW ZLOO FROOLGH ZLWK D URERW RI KLJKHU SULRULW\ DORQJ LWV SDWK ,Q WKLV H[SHULPHQW ZH DSSUR[LPDWH WKLV SUREDELOLW\ PRUH GLUHFWO\ WKDQ ZH GLG IRU WKH ÀUVW VFHQDULR /HW XV DVVXPH ZH DUH SODQQLQJ IRU URERW j DQG WKDW D SDWK KDV DOUHDG\ EHHQ SODQQHG IRU URERWV 1, . . . , j−1 $V WKH URERWV DUH GLVFVKDSHG RQO\ WKHLU SRVLWLRQ LQÁXHQFHV ZKHWKHU RU QRW WKH\ FROOLGH /HW N (pit , Σit ) EH WKH PDUJLQDO SUREDELOLW\ GLVWULEXWLRQ RI WKH SRVLWLRQ RI URERW i DW VWDJH t DORQJ i·V SDWK DV FRPSXWHG E\ /4*03 7KHQ WKH GLVWULEXWLRQ RI WKH UHODWLYH SRVLWLRQ RI URERW j DQG URERW i IRU i ∈ 1, . . . , j − 1 DW VWDJH t LV N (pit − pjt , Σit + Σjt ) 7KH SUREDELOLW\ Pt (i ⊗ j) WKDW URERW j FROOLGHV ZLWK URERW i DW VWDJH t LV WKHQ JLYHQ E\  j −1 T i (p−pij exp(− 12 (p−pij t ) (Σt +Σt ) t )) dp,  2π det(Σit + Σjt )1/2 p@ 5 $OWHURYLW] 7 6LPHRQ . *ROGEHUJ 7KH VWRFKDVWLF PRWLRQ URDG PDS D VDPSOLQJ IUDPHZRUN IRU SODQQLQJ ZLWK 0DUNRY PRWLRQ XQFHU WDLQW\ 3URF 5RERWLFV 6FLHQFH DQG 6\VWHPV  >@ - YDQ GHQ %HUJ 0 2YHUPDUV 3ULRULWL]HG PRWLRQ SODQQLQJ IRU PXOWLSOH URERWV 3URF ,(((56- ,QW &RQI RQ ,QWHOOLJHQW 5RERWV DQG 6\VWHPV  >@ * YDQ GHQ %HUJHQ &ROOLVLRQ GHWHFWLRQ LQ LQWHUDFWLYH ' HQYLURQPHQWV 0RUJDQ .DXIPDQQ 3XEOLVKHUV 



136

>@ ' %HUWVHNDV '\QDPLF SURJUDPPLQJ DQG RSWLPDO FRQWURO $WKHQD 6FLHQWLÀ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ÀFLDO ,QWHOOLJHQFH   ²  >@ / .DYUDNL 3 6YHVWND -& /DWRPEH 0 2YHUPDUV 3UREDELOLVWLF URDGPDSV IRU SDWK SODQQLQJ LQ KLJK GLPHQVLRQDO FRQÀJXUDWLRQ VSDFHV ,((( 7UDQV RQ 5RERWLFV DQG $XWRPDWLRQ ²   >@ * .HZODQL * ,VKLJDPL . ,DJQHPPD 6WRFKDVWLF PRELOLW\EDVHG SDWK SODQQLQJ LQ XQFHUWDLQ HQYLURQPHQWV 3URF ,(((56- ,QW &RQI RQ ,QWHOOLJHQW 5RERWV DQG 6\VWHPV  >@ @ + .XUQLDZDWL ' +VX : /HH 6$5623 (IÀFLHQW SRLQWEDVHG 320'3 SODQQLQJ E\ DSSUR[LPDWLQJ RSWLPDOO\ UHDFKDEOH EHOLHI VSDFHV 3URF 5RERWLFV 6FLHQFH DQG 6\VWHPV  >@ 6 /D9DOOH 6 +XWFKLQVRQ $Q REMHFWLYHEDVHG IUDPHZRUN IRU PRWLRQ SODQQLQJ XQGHU VHQVLQJ DQG FRQWURO XQFHUWDLQWLHV ,QW - RI 5RERWLFV 5HVHDUFK  ²  >@ 6 /D9DOOH - .XIIQHU 5DQGRPL]HG NLQRG\QDPLF SODQQLQJ ,QW -RXU QDO RQ 5RERWLFV 5HVHDUFK  ²  >@ 6 /D9DOOH 3ODQQLQJ DOJRULWKPV &DPEULGJH 8QLYHUVLW\ 3UHVV  >@ $ /D]DQDV - /DWRPEH 0RWLRQ SODQQLQJ ZLWK XQFHUWDLQW\ D ODQGPDUN DSSURDFK $UWLÀFLDO ,QWHOOLJHQFH  ²  >@ 1 0HOFKLRU 5 6LPPRQV 3DUWLFOH 557 IRU SDWK SODQQLQJ ZLWK XQFHUWDLQW\ 3URF ,((( ,QW &RQI RQ 5RERWLFV DQG $XWRPDWLRQ  >@ 3 0LVVLXUR 1 5R\ $GDSWLQJ SUREDELOLVWLF URDGPDSV WR KDQGOH XQFHUWDLQ PDSV 3URF ,((( ,QW &RQI RQ 5RERWLFV DQG $XWRPDWLRQ  >@ $ 1DNKDHL ) /DPLUDX[ $ IUDPHZRUN IRU SODQQLQJ PRWLRQV LQ VWRFKDVWLF PDSV 3URF ,QW &RQI RQ &RQWURO $XWRPDWLRQ 5RERWLFV DQG 9LVLRQ  >@ & 3DSDGLPLWULRX - 7VLVLNOLV 7KH FRPSOH[LW\ RI 0DUNRY GHFLVLRQ SURFHVVHV 0DWKHPDWLFV RI 2SHUDWLRQV 5HVHDUFK  ²  >@ 5 3HS\ $ /DPEHUW 6DIH SDWK SODQQLQJ LQ DQ XQFHUWDLQFRQÀJXUDWLRQ VSDFH XVLQJ 557 3URF ,(((56- ,QW &RQI RQ ,QWHOOLJHQW 5RERWV DQG 6\VWHPV  >@ - 3RUWD 1 9ODVVLV 0 6SDDQ 3 3RXSDUW 3RLQWEDVHG YDOXH LWHUDWLRQ IRU FRQWLQXRXV 320'3V -RXUQDO RI 0DFKLQH /HDUQLQJ 5HVHDUFK   >@ 6 3UHQWLFH 1 5R\ 7KH EHOLHI URDGPDS HIÀFLHQW SODQQLQJ LQ OLQHDU 320'3V E\ IDFWRULQJ WKH FRYDULDQFH 3URF ,QW 6\PS RI 5RERWLFV 5HVHDUFK  >@ 1 5R\ : %XUJDUG ' )R[ 6 7KUXQ &RDVWDO QDYLJDWLRQ  PRELOH URERW QDYLJDWLRQ ZLWK XQFHUWDLQW\ LQ G\QDPLF HQYLURQPHQWV 3URF ,((( ,QW &RQI RQ 5RERWLFV DQG $XWRPDWLRQ  >@ 5 7HGUDNH /45WUHHV )HHGEDFN PRWLRQ SODQQLQJ RQ VSDUVH UDQGRP L]HG WUHHV 3URF 5RERWLFV 6FLHQFH DQG 6\VWHPV  >@ 6 7KUXQ : %XUJDUG ' )R[ 3UREDELOLVWLF 5RERWLFV 0,7 3UHVV  >@ * :HOFK * %LVKRS $Q LQWURGXFWLRQ WR WKH .DOPDQ ÀOWHU 7HFK 5HSRUW 75  8QLYHUVLW\ RI 1RUWK &DUROLQD DW &KDSHO +LOO  >@ :LNLSHGLD &KLVTXDUH GLVWULEXWLRQ KWWSHQZLNLSHGLDRUJZLNL&KL VTXDUH 

The Smooth Curvature Flexure Model: An Accurate, Low-dimensional Approach for Robot Analysis Lael U. Odhner and Aaron M. Dollar Yale University Email: {lael.odhner, aaron.dollar}@yale.edu Abstract— This paper presents a new and comprehensive method for modeling robots having highly flexible members such as flexure joints. An accurate model of large deformation bending is important for precisely describing the configuration of the flexible member. Additionally, the accuracy of the Jacobian and Hessian of the forward kinematics are critically important at large angles for predicting the deformation and the stiffness of the joint under load. The model introduced here is based on the assumption that the curvature of a beam in bending is smooth, and thus can be approximated by low-order polynomials. This produces a parameterized description of flexure motion that can be used as a joint model when expressed in Denavit-Hartenberg form, as a transformation from one rigid link to the next in a serial manipulator. We will show that with only three parameters, this model faithfully reproduces the elastic deformation of a flexure hinge predicted by the continuum model, even for large angles, without requiring numerical integration or many finite elements. It can also be used to compute the compressive buckling load of the flexure as predicted by the continuum model.

I.

INTRODUCTION

Highly flexible members have been frequently considered in the context of robotic hardware. A number of studies into the behavior of flexible link robots have been conducted, often for the purposes of controlling for undesirable dynamic effects related to working with long, thin links (e.g. [1-3]). A smaller number of efforts have dealt with the beneficial aspects of highly flexible links, such as providing a large number of degrees of freedom for manipulation tasks [4] or low stiffness for grasping and assembly purposes [5]. A related application is the use of highly flexible members as joints between rigid links, typically referred to as flexures. The contrast between these two applications is shown in Fig. 1. These features are commonly used to allow motion in monolithic structures, and have been used as joints in a number of different robotic mechanisms, particularly in compliant hands [6-8]. The benefits of flexure-based joints include having no sliding parts (and therefore no friction or stick-slip effects), no backlash, and are able to compliantly deform in response to unplanned collisions, making them ideal for robots that must operate in unstructured environments [9]. Another major benefit of flexures is the simplicity and potential lower cost compared to standard revolute joints, which require bearings for smooth, accurate motion. As robots become more common as commercial products, flexures are

Rigid Links Pin Joints

Flexible Links Pin Joints

Rigid Links Flexure Joints

Figure 1. A comparison between traditional rigid robot manipulators (left), flexible link robots (center), and flexure-joint robots, in which flexible links act as hinges (right).

likely to be used with increasing frequency due to their compatibility with inexpensive polymeric fabrication processes such as multi-shot injection molding [10] and shape deposition manufacturing [11]. One drawback to flexure-based robot mechanisms is the complex mechanical behavior that they exhibit compared to pin joints. A pin joint has one degree of freedom, whereas the elastic deformation of a beam in bending has infinitely many degrees of freedom. Moreover, a flexure hinge in a robot often bends to angles up to 90 degrees or more, so classical smalldeflection beam bending models are inapplicable. As a result, there is no canonical parametric model for planar flexure hinges suitable for robot analysis. This paper presents a model that can fill this role. In order to apply the rich set of tools available for serial manipulator design, control, and analysis, one must have a model of elastic behavior that is accurate and computationally simple. The design specification for a good flexure model can be broken into three functional requirements: 1. It should be possible to compute both the shape of the robot and the elastic energy associated with deformation as a function of a small set of generalized coordinates, as one might describe a jointed mechanism using the internal joint angles. 2. It should be possible to compute the force in generalized coordinates resulting from a force on the robot at any point using the Jacobian of that point’s coordinates, as well as the local equilibrium position resulting from such a force. 3. It should be possible to compute the stiffness in generalized coordinates resulting from a force on the robot at a point by the Hessian of that point’s coordinates, as well as any buckling modes the robot has (configurations/loads having zero stiffness in some direction).

 137

Many models of flexible robot components meet some, but not all of these requirements. One common approach is to model flexure elements as having constant curvature [12]. Flexures have also been approximated as a single pin joint halfway between the ends of the flexure [13]. Both of these models capture the relative rotation between rigid bodies on the manipulator, making them useful for inverse kinematic computation and form closure grasp analysis. However, because these models have fewer degrees of freedom than a real flexure, they are too rigid and under-predict the deformation of a loaded flexure. Another family of models, called pseudo-rigid body models, consists of one or several joints placed to approximate the flexure’s center of rotation, connected in parallel with nonlinear springs fit by regression to the exact force-deflection profile [14]. These models can be used to find flexure deflection under load, but they are unsuitable for manipulator analysis because the linkage geometries used to approximate the beam bending change based on the direction of applied force, and thus are not purely kinematic descriptions of flexure behavior. Another approach is based on assuming some set of superimposed deformational modes [15]. This technique has been applied to flexible links (e.g. [16]) and continuum manipulators [17]. Modal models for flexures have been proposed based on analytically calculated small-deformation solutions, as well as finite element solutions [18,19]. However, modal models are specific to the behavior they are designed to model, and none of the currently available models in the robotics literature accurately capture largedeformation flexure behavior. The flexure model discussed in this paper is a modal model which approximates the curvature of a flexure using a polynomial basis. The polynomial coefficients define the relative position and orientation of two bodies connected by the flexure, as well as the elastic energy stored in the flexure itself. This model meets all three of the functional requirements introduced above, while avoiding the need to use numerical integration or to break the beam into many finite elements. It predicts not only the deflection of a flexure under load, but also second-order kinematic effects such as buckling and the change in flexure stiffness resulting from compressive or tensile loads. These second-order effects are particularly useful in the study of grasping and manipulation, where grasp stability may depend on the elastic stability of the manipulator itself [20]. Previous study of this model has dealt with stiffness prediction [21]. This paper is intended as a more general discussion of this model’s applicability to robotic systems. The remainder of this paper is divided into three sections. Section II is an overview of the flexure model, describing how the parameters define the shape and the energy function of a flexure hinge. Section III examines the shape of the flexure when an arbitrary load and moment is applied at one end. The results are compared to exact large deformation beam solutions. Additional results are shown comparing the smooth curvature model to finite element models of a sample mechanism. Section IV demonstrates the second-order kinematic accuracy of the model by comparing classical continuum buckling models to the discrete buckling predicted by the proposed parameterized model.

Figure 2. A comparison of the moment profile in small- and largedeflection cantilevered beam bending.

II.

THE SMOOTH CURVATURE MODEL

A. Motivation In 1694, Jacob Bernoulli proposed (and solved) the problem of finding the shape of a pre-bent cantilevered beam of length L that would bend into a straight line when loaded with an arbitrarily large force P at the tip [22]. Today the curve is known as the clothoid or Euler spiral. The curvature κ(s) of the clothoid curve varies linearly with the arc length s from the base of the cantilever,

κ (s) =

P ⎛ s⎞ ⎜1 −  EI ⎝ L ⎠

(1)

Here E is the elastic modulus and I is the constant planar moment of the beam area. The clothoid is also a passable approximate solution the more useful problem of finding the deformed shape of an initially straight beam loaded at one end with a large load. This can be seen by examining the nonlinear deformation of a cantilevered beam, as shown in Fig. 2. For a small end load (left), the bending moment will be almost exactly proportional to the distance from the tip of the flexure, as expected. A large end load (right) will produce a non-linear deformation profile, but the moment, plotted as a function of arc length, is still roughly linear. The curvature is directly proportional to bending moment in the beam, τ(s),

κ (s) =

τ ( s) EI

(2)

For some range of large loads, then, the curvature of a beam can be approximated as some constant plus a linear function of arc length. Horn discussed this approximation in the context of spline curves [23]. The accuracy of this model could be further improved by noting that while the curvature may not be exactly linear, it is certainly smooth, and so might be described with a basis of n Legendre polynomials, G0(s)…Gn-1(s). The curvature is expressed as a weighted sum of the bases,

κ ( s, q) = q 0 G 0 ( s) + q1 G1 ( s) + K + q n −1 G n −1 ( s)

(3)

The coefficients, q0…qn-1, used as a generalized coordinate vector q for describing the deformation of the flexure, are central to the proposed flexure model. Two particular cases will be considered here, corresponding to the models with 2

 138

and 3 parameters, whose basis functions are Legendre polynomials, translated and scaled to be orthogonal on [0, L],

κ 2 ( s, q ) = q 0

1 2s − L + q1 L L2

(4) 2

κ 3 ( s, q ) = q 0

1 2s − L 6s − 6sL + L + q1 + q2 L L2 L3

Proximal Coordinate Frame

2

(5)

Models of this type, which we will call smooth curvature models, can be used to predict the shape of the flexure in bending, as well as the elastic bending energy. These derivations follow in the next two sections.

Figure 3. A kinematic model of a flexure consists of a transformation mapping the coordinate frame at the proximal side of the joint to the coordinate frame at the distal side of the joint.

B. Flexure Shape Joints in a serial robot manipulator are often represented in Denavit-Hartenberg notation, that is, as a parameterized axial, radial and angular transformation from one joint to another. The analogous transformation for a pin joint is a rotation about the joint axis. In the case of a flexure, this transformation corresponds to the translation and rotation from one end of the flexure to the other, as shown in Fig. 3 [12]. This could be written as a matrix, for instance,

⎡cos(ϕ tip ) − sin(ϕ tip ) T = ⎢⎢ sin(ϕ tip ) cos(ϕ tip ) ⎢⎣ 0 0

x tip ⎤ y tip  1 ⎦

(6)

The three quantities characterizing this transformation are the flexure tip displacement (xtip, ytip), and the relative angle from the base of the flexure to the tip of the flexure, φtip. They can be written as functions of the polynomial coefficient vector, q. The angular profile, φ(s, q), is the integral of the curvature,

ϕ ( s, q) = ∫ κ ( s, q )ds ϕ 2 ( s, q ) = q 0

s s 2 − sL + q1 L L2

s 2 − sL s 2 s 3 − 3s 2 L + sL2 ϕ 3 ( s, q ) = q 0 + q 1 + q2 2 L L L3

(7)

L

(11)

0 L

y tip = ∫ sin(ϕ ( s, q ))ds

L

U (q) =

EI κ ( s, q) 2 ds 2 ∫0

(12)

0

These expressions are transcendental. The second order solution can be solved in terms of Fresnel integrals by

(13)

If the 2 parameter curvature is used, U(q) evaluates to a weighted sum of the squared parameters.

EI 2

2

⎛ q 0 4 q1 ( 2 s − L ) ⎞ EI ⎛ 2 q12 ⎞ ⎜ q0 +  (14) ⎜⎜ +  = ds  L L2 2 L ⎜⎝ 3 ⎠ ⎠ 0 ⎝

L

∫

Because the polynomial basis is orthogonal under convolution over the interval [0, L], there are no cross-terms in this expression. The expression for energy given 3 parameters differs only in the addition of a single term:

U 3 (q) =

(10)

The tip position of the flexure can be found by integrating the cosine and sine of the angular profile,

x tip = ∫ cos(ϕ ( s, q ))ds

C. Elastic Energy Having found the shape of the hinge as a function of q, we now turn to finding an expression for the elastic energy in the flexure. The energy stored in an Euler-Bernoulli beam is proportional to the integral of the squared curvature [15]:

U 2 (q) =

At the end of the flexure (s=L), the tip angle is equivalent to q0 irrespective of the model order. This is a happy side effect of using orthogonal polynomials, as all the higher, non-constant terms must integrate to zero:

ϕ tip = ϕ ( L, q) = q 0

completing the square and using trigonometric addition identities, but it contains discontinuities, and is not practically useful. Further, this strategy does not generalize to the 3 parameter model. Instead, an interpolating approximation was used, so that (11) and (12) can be analytically approximated within some reliable error bounds. In this paper, Chebyshev interpolation [24] was used to approximate the sine and cosine functions as polynomials. There is a trade-off between the domain of interpolation and the computational cost, so maximum flexure rotation was limited to be less than 108°. Alternatively, Gaussian quadrature could be used, essentially interpolating the entire integrand as a polynomial.

(8) (9)

Distal Coordinate Frame

EI ⎛ 2 q12 q 22 ⎞ ⎜ q0 +  + 2 L ⎜⎝ 3 5 ⎠

(15)

D. Summary The smooth curvature model for flexure hinges has been introduced, based on the observation that the curvature of a flexure can be approximated using a low-dimensional basis of orthogonal polynomials. The position and orientation of the flexure tip can be found relative to its base using only a few model parameters, and these can be used to represent the flexure as a joint in Denavit-Hartenberg form. The elastic energy stored in the bent beam is a weighted sum of the squared flexure parameters.

 139

The remaining sections will demonstrate that this model can be used to satisfy the two other requirements for a flexure joint model, that is, that the model accurately predicts the equilibrium position of the flexure when an arbitrary force and moment are applied, and that the model accurately predicts variable stiffness effects and buckling due to compressive loads. III.

DEFLECTION UNDER LOAD

A. Jacobian Analysis of the Forces on a Robot Accurate descriptions of the force exerted on a robot and the resulting deflection are central to many problems in the control and analysis of robot manipulators. The net generalized force F on a manipulator experiencing a force fp at some point p is given by the Jacobian of that point’s coordinates, and the gradient of the potential energy function U(q),

F = ∇ q ( p) T f

p

+ ∇ q (U (q ))

(16)

Figure 4. A flexure, loaded at the end by a force and a moment. This figure shows the direction of load, θ.

deformation beam bending 1 . The equilibrium configurations were compared to the exact solution obtained by numerically integrating the large-deformation Euler-Bernoulli equation,

~ s ) ⎤ ⎡ P (cos(θ ) cos(ϕ (~ ⎡τ~ (~ s )) − sin(θ ) sin(ϕ (~ s )))⎤ ⎢   ⎢ϕ (~ ~ ~ τ (s ) d ⎢ s ) ⎢  =  x (~ s ) ⎢ d~ s ⎢~ cos(ϕ (~ s ))  ⎢~ ~  ⎢ ~ sin(ϕ ( s )) ⎣ y ( s )⎦ ⎣⎢ ⎦

The generalized force balance equation is only realistic if the generalized coordinates faithfully represent all of the motions that the robot is capable of making. For a flexure hinge, it is important that the motion of flexure tip, as described in the previous section, is accurate, so that the forces and moments transmitted from one link to the next result in a physically realistic deformation of the flexure hinges.

This is a restatement of (2), (7), (11) and (12) in differential form, after applying the substitutions from (17). This equation was solved using a Runge-Kutta solver. The tip moment M, force angle θ, and tip angle φtip, were specified, and the integral from the tip of the flexure to the based was computed. The tip force, P, was found using a bisection search such that the boundary conditions at both ends of the flexure were simultaneously satisfied.

This section considers two tests for benchmarking the ability of the smooth curvature flexure model to predict deformation under load. The first test compares the exact deflection of a cantilevered flexure (using numerically computed elastica curves [25]) to the tip position predicted by the smooth curvature model. The second test computes the deflection of a two-link finger from a tendon-driven elastic gripper developed by the authors. The two flexure hinges in the finger are modeled with finite element flexure models, and with the 3 parameter smooth curvature model, showing that the two models agree despite the vastly reduced parameter space of the smooth curvature model.

Each force-moment combination (P, θ, M) that was computed for the exact beam equation was applied to the tip of the smooth curvature flexure model, using the generalized force balance from (16). The Jacobian of the tip coordinates xtip, ytip and φtip was derived from (10), (11) and (12),

⎧1, i = 0 =⎨ ∂q i 0, i ≠ 0

∂ϕ tip

B. Tip Deflection of a Loaded Flexure A straightforward method of examining the accuracy of a flexible beam is to clamp one end and examine the deflection of the other end when subject to an arbitrary moment M and force P exerted at an angle θ, as shown in Fig. 4. To ensure proper scaling of the results, a non-dimensional form of the beam bending equations should be used, based on these substitutions:

~ PL2 ~ ML ~ τL ~ s ~ x ~ y P= ,M = ,τ = ,s = ,x= ,y= L L L EI EI EI

(18)

(17)

The dimensionless beam bending equations are equivalent to modeling a beam as having length 1, and an elastic modulus and cross-sectional moment equal to 1.The only parameter that is unaffected by this scaling is the beam’s angular profile, φ(s). Results will be computed for the case when the tip angle, φtip, is equal to 90°, a prototypical test case in the study of large-

∂x tip ∂q i ∂y tip ∂q i

(19)

L

∫

= − sin(ϕ ( s, q)) 0 L

∫

= cos(ϕ ( s, q )) 0

∂ϕ tip ∂q i

∂ϕ tip ∂q i

ds

ds

(20)

(21)

As in (11) and (12), Chebyshev interpolation was used to produce analytical approximations of (20) and (21). The derivatives of the energy function are much simpler, and can be found from (15),

1 This is a generalization of the rectangular elastica problem posed by Bernoulli, the problem of finding the shape of a cantilevered beam bent at a right angle by a force at the tip [25].

 140

0 ⎤ ⎡q 0 ⎤ ⎡1 0 EI ⎢ 0 1 / 3 0  ⎢⎢ q1  ∇ qU (q) = ⎢ L ⎢⎣0 0 1 / 5⎦ ⎢⎣q 2 ⎦

(22)

The force balance, computed from (19)-(22), was set to zero to form a system of nonlinear equations, which was solved numerically. Three force-moment combinations were used to compare the two models. These combinations, labeled A, B, and C, are shown in Fig. 5. In case A, a pure bending moment was applied, sufficient to bend the flexure to an angle of 90°. Case B was a pure force of a magnitude sufficient to bend the flexure to 90°. Case C was the most complex load, consisting of a moment equal and opposite to the moment applied in combination 1, counteracted by a force so that the flexure tip angle remained at 90°. The prediction errors from the smooth curvature models were computed for values of θ ranging from 20 degrees to 105 degrees (as in Fig. 4). The lower bound of 20 degrees was chosen because the magnitude of the load required to bend a beam to 90° in case B has a vertical asymptote at θ=0. Thus, the behavior of the flexure becomes increasingly unrealistic in this case. The upper bound of 105 degrees was chosen because the elastica curves generated with the Runge-Kutta solver could not predict buckled configurations, and as the force on the flexure tip becomes increasingly compressive (i.e. θ > 90°), good reference comparisons could not be made. Instead, compressive loads were compared to finite element models in the following subsection. The positional error of the flexure tip was found, that is, the norm of the vector from the predicted flexure tip to the tip of the numerically computed elastica curve. This is plotted in nondimensional form, meaning that the error is given as a fraction of the flexure length. The angular error is also shown. The errors for the 2 parameter model are shown in Fig. 6, and the errors for the 3 parameter model are shown in Fig. 7. These plots show a number of significant results. First, the errors in case A (pure moment loading) were very small for both the 2 and 3 parameter smooth curvature models. The exact shape one would expect for a beam having a constant bending moment is an arc, a shape that can be exactly reproduced with both models. Thus, the error is correspondingly small. The errors observed in loading cases B and C indicate that the 3 parameter model is significantly more accurate, especially when the flexure is loaded by an opposed force and moment. Most importantly, in every case, the 3 parameter smooth curvature model was within a positional accuracy of 1 percent of the beam length, and an angular accuracy bound of 1°. C. Finite Element Comparisons One purpose of the smooth curvature model of particular interest to the authors is to enable efficient analysis of manipulators having multiple flexure joints. In previous work, the authors have developed robot hands incorporating polymeric elastic flexure joints [8,9]. These hands are made up of 2-link, tendon-driven fingers, represented in Fig. 8A. In order to evaluate the usefulness of the smooth curvature model for multi-link manipulators, a finite element model was constructed using an object-oriented Matlab library created by

 141

Figure 5. The three loading cases used to test the smooth curvature model.

Figure 6. Prediction errors for the 2 parameter smooth curvature model. Cases A, B, and C correspond to the cases in Fig. 5.

Figure 7. Prediction errors for the 3 parameter smooth curvature model. Cases A, B, and C correspond to the cases in Fig. 5.

TABLE I.

FEM VS. SMOOTH CURVATURE FINGER MODEL RESULTS

Case FEM, Tendon Force (i) SC, Tendon Force (ii) FEM, Tendon + Pad Force (iii) SC, Tendon + Pad Force (iv)

Pad x -0.0048 -0.0047 0.0615 0.0615

Pad y 0.1012 0.1012 0.1046 0.1046

Pad angle 147.6553 147.5760 103.0476 103.0168

A. Flexure-based robot finger Tendon

Smooth Curvature Model

B. Tendon force

Finite Element Model

C. Tendon force + pad force Figure 9. A single tendon-driven joint, modeled using the smooth curvature model (left) and finite rotational elements (right).

Figure 8. A comparison between FEM and smooth curvature (SC) models: A. The FEM model, no tendon force vs 10 N tendon force. B. SC model (i) vs FEM (ii), 10 N tendon force. C. SC model (iii) vs. FEM (iv), 10 N force + 0.5 N force on distal link.

the authors [26]. This model represents each flexure in a fashion similar to the finite element model proposed in [1], consisting of 16 small rotational links. The transformation representing each joint in Denavit-Hartenberg form was the composition of the many resulting rotations and translations making up each finite element. A model of the finger was also constructed using our 3 parameter smooth curvature flexures. This model uses the joint transformation from (6) and the energy function from (15) to describe the joint behavior in terms of three generalized coordinates per joint, for a total of 6. The finite element model, by comparison, had a total of 32 generalized coordinates. Both models were subjected to two different loading conditions: in the first condition, the finger was actuated with a single tendon connected to the distal link, as in Fig. 8B. The second condition, depicted in Fig. 8C, included the same tendon force and a horizontal force applied to the center of the pad on the distal link. In both cases, the generalized force balance was computed using (16), computing the kinematics of the tendon and the center of the distal pad using a composition of geometric joint and link transformations. The results of the test are shown in Fig. 8 and Table I, which describes the agreement between the smooth curvature and finite element models as to the position and orientation of the center of the distal link. The results show that the displacement of the distal link pad as calculated by the smooth curvature model is within 0.1% of the position predicted by the finite element model. The angular agreement is similarly within a tenth of a degree in both cases. Attempts at visual comparison between the FEM and smooth curvature models by overlaying the two were unsuccessful, because they were almost indistinguishable to the eye.

D. Summary For both a single flexure hinge and a two-joint manipulator, the smooth curvature model has been shown to accurately predict the deformation of a manipulator under a wide range of forces and moments. This is interesting and new because most methods of accurately solving large-deflection beam bending problems involve numerical integration, or the breaking down of a beam into many finite elements. Because the smooth curvature model achieves a useful degree of accuracy with only three parameters per joint, calculation of dynamics and statics for control or motion planning is a much simpler process. IV.

STIFFNESS AND BUCKLING

A. Stiffness of a Loaded Elastic Structure One major difference between the well-studied problem of flexible-link manipulators and the newer field of flexurejointed manipulators is the relatively increased importance of buckling in flexure joints. The flexure joints shown in Fig. 9 show a typical tendon/flexure actuation scheme. The tendon exerts a force in tension, which is balanced by a compressive force in the flexure, unless a parallel load path exists. This compressive force is quite large, and can easily approach the Euler buckling load of a thin flexure (the load at which the lateral stiffness of the flexure is zero). Unlike structural columns, buckling does not represent a necessarily undesirable effect. The fact that the flexure is buckled just means that its rotational stiffness is very low. Most pin joints, for example, have zero rotational stiffness and this is not an impediment to their use in robots. However, this change in stiffness as a function of load is critically important in some robotic tasks. For example, a change in joint stiffness will affect computed torque control models. Additionally, some tasks such as grasping and manipulation rely on the elastic stability of the

 142

whole system; a buckling mode could be harmless, or it could correspond to configuration in which a grasped object twists out of its gripped position [20]. As a result of all these concerns, it is important that a flexure model provide a reasonable model of elastic buckling. This section briefly describes the ability of the smooth curvature model to predict elastic buckling in a flexure using the Hessian of the flexure kinematics. As a proxy case for comparing the continuum behavior of a flexure to the smooth curvature model, the smooth curvature model will be used to predict buckling in compression by finding the smallest compressive load for which the generalized stiffness matrix is singular. This result can be compared to Euler’s buckling load formula. The 2 and 3 parameter models produce successively better approximations of buckling.

These are polynomials, which can be evaluated to compute the Hessian of xtip with respect to the generalized coordinates. For the 2 parameter model, the Hessian is a 2 by 2 matrix,

⎡ 1 / 3 − 1 / 12⎤ ∇ q ∇ q ( x tip ) = L ⎢  ⎣− 1 / 12 1 / 30 ⎦

The stiffness due to potential energy can be found by taking the Hessian of the energy function derived in (14),

∇ q ∇ q (U (q)) =

(23)

∂ 2 x tip ∂q i ∂q j

L

= −∫ 0

∂φ ∂φ ds ∂q i ∂q j

(26)

(29)

1 12 EI 1 − 2 3P2 L 30

=0

(30)

(31)

This is only 0.75% larger than the true value reported by the continuum model in (23). This exercise can be repeated for the three parameter model, to find the predicted buckling load, P3,

EI 1 − 2 P3 L 3 1 12 1 60

1 12

1 60

EI 1 − 2 3P3 L 30

0

0

EI 1 − 5 P3 L2 210

=0

(32)

The resulting buckling load prediction is within 0.02% of the value predicted by Euler’s beam buckling formula,

(25)

When the flexure is straight, q0=q1=q2=0. In this configuration, (25) can be simplified, because the sine term disappears and the cosine term approaches one,

⎡1 0 ⎤ ⎡ 1 / 3 − 1 / 12⎤ ⎢  − PL ⎢  ⎣0 1 / 3⎦ ⎣− 1 / 12 1 / 30 ⎦

P2 L2 = 2.4860... EI

Stiffness is a function of the Hessian of the coordinates where force is applied, and of the Hessian of the energy, U(q). When this stiffness matrix has an eigenvalue that is zero or negative, it buckles. In other words, there will exist some eigenvector δq, which, when applied to the robot as a perturbation, will produce a destabilizing force. This could also be thought of as a test for the convexity of the total energy in the robot.

L ⎡ ∂ 2 xtip ∂φ ∂φ ∂ 2φ ⎤ = − ∫ ⎢cos(φ ( s)) + sin(φ ( s))  ds ∂qi ∂q j ∂qi ∂q j ∂qi ∂q j ⎦ 0 ⎢ ⎣

(28)

The smallest root of this polynomial is the most physically meaningful, as it represents the load at which the unconstrained flexure will buckle,

(24)

We will derive the generalized stiffness matrix for the smooth curvature model, when the flexure is loaded in the -x direction with a force, P. Thus, the contact point p from (24) above is the scalar xtip, as described in (8). The Hessian elements can be calculated from the Jacobian of xtip in (20),

EI L

EI 1 − P2 L2 3 1 12

The concept analogous to elastic buckling in a generalized coordinate model has to do with the generalized stiffness matrix obtained by taking the gradient of the generalized force balance from (16) with respect to q,

K = ∇ q ∇ q ( p ) T f p + ∇ q ∇ q (U (q ))

⎡1 0 ⎤ ⎢  ⎣0 1 / 3⎦

The buckling load of the 2 parameter model, P2, is the value of P for which the determinant of K is zero, indicating that the matrix has an eigenvalue equal to zero,

2

Pcrit L2 ⎛ π ⎞ = ⎜  = 2.4674... EI ⎝2⎠

EI L

The coordinate Hessian and the energy Hessian can be substituted back into (24),

K=

B. Continuum vs. Discrete Buckling A continuum structure is said to buckle when it has zero (or negative) stiffness in some direction, so that a small perturbation to the structure’s shape is met by a destabilizing force, rather than a restoring force. The compressive load Pcrit at which a clamped-free beam should buckle is given by Euler’s well-known formula [27]:

(27)

P3 L2 = 2.4677... EI

(33)

C. Discussion The results of this brief study indicate that the smooth curvature model has no difficulty predicting the stiffness of a straightened flexure hinge as a function of load. A more general argument, too long to be presented here, would also cover the stiffness of a constrained flexure, such as one which is pinned or clamped at the distal end. Although the effect of constraints on the generalized stiffness is lengthier to describe,

 143

[5]

the 3 parameter model can successfully describe these cases. Finally, it is worth mentioning that although stiffness prediction of a flexure undergoing large deformation has not been presented here, the smooth curvature model will predict the stiffness matrix for large deformations, as we have shown in [21]. V.

[6]

[7]

CONCLUSIONS

A. In Summary In this paper, we have presented a model for flexible links that is accurate for large deformations, so that it can be used for the special case of flexible links as flexure hinges. These models are compatible with all of the standard tools used for manipulator analysis, because they are in a form where the shape of the joint and the elastic energy of the joint can be entirely described by a set of generalized coordinates. We have demonstrated that a flexure can be described to a high level of accuracy using only three parameters – arguably the minimum number of parameters capable of describing a flexure with three independent end conditions, xtip, ytip and φtip. This model is useful for “zeroth” order descriptions (shape and energy), first order descriptions (local deformation and force), and for second order descriptions (buckling configurations and stiffness) of mechanical behavior of flexible members undergoing large deformations under loads.

[8]

[9]

[10]

[11]

[12]

[13]

[14]

B. Future Directions The smooth curvature flexure model presents several obvious directions for further study. First, this model does not take into account axial deformation, which is important in accurately analyzing parallel mechanisms. We are confident that this method can be extended to describe such mechanisms through the addition of one or two axial deformation modes. Second, it is worth noting that although models of planar beam bending are quite useful, many flexures admit a great deal of out-of-plane motion. This behavior can be characterized by modal models similar to the one presented here. We are working on extensions of smooth curvature models to three dimensions, including possible techniques for coping with the greatly increased complexity of describing non-commutative spatial rotations. REFERENCES [1]

[2]

[3]

[4]

E. Bayo, “A finite-element approach to control the end-point motion of a single-link flexible robot,” Journal of Robotic Systems, vol. 4(1), pp. 6375, 1986. H. Asada, Z.-D. Ma, and H. Tokumaru, “Inverse Dynamics of Flexible Robot Arms: Modeling and Computation for Trajectory Control”, J. Dynamic Systems, Measurement, and Control vol.112(2), pp. 177-186, 1990. Zheng-Hua Luo, “Direct strain feedback control of flexible robot arms: new theoretical and experimental results,” IEEE Transactions on Automatic Control, vol.38(11), pp.1610-1622, 1993. A. Hemami, “Studies on a light weight and flexible robot manipulator,” Robotics, vol. 1(1), pp. 27-36, 1985.

[15] [16]

[17]

[18]

[19]

[20]

[21]

[22] [23] [24] [25] [26] [27]

 144

S. Sur and R.M. Murray, “An Experimental Comparison of Tradeoffs in Using Compliant Manipulators for Robotic Grasping Tasks,”Proceedings of the 1997 IEEE International Conference on Robotics and Automation, pp. 1807-1814, 1997. F. Lotti and G. Vassura, “A Novel Approach to Mechanical Design of Articulated Fingers for Robotic Hands,” Proceedings of the 2002 IEEE International Conference on Intelligent Robots and Systems, pp. 16871692, 2002. F. Lotti, P. Tiezzi, G. Vassura, L. Biagiotti, G. Palli, C. Melchiorri, “Development of UB Hand 3: Early Results,” Proceedings of the 2005 IEEE International Conference on Robotics and Automation, pp. 4499 – 4504, 2005. A. Dollar and R. Howe, “The Highly Adaptive SDM Hand: Design and Performance Evaluation,” Int. Journal of Robotics Research, v.29, pp. 585-597, 2010. A. Dollar and R. Howe, “Towards grasping in unstructured environments: Grasper compliance and configuration optimization,” Advanced Robotics, v. 19, n. 5, pp. 523-543, 2005. R. Digiantonio and G. Lawrence, “Two-shot molding of thermoplastic elastomers,” Proceedings of the 1992 SPE Annual Technology Conference, pp. 851-853, 1992 R. Merz, F. Prinz, K. Ramaswami, M. Terk, L. Weiss, “Shape deposition manufacturing,” Proceedings of the Solid Freeform Fabrication Symposium, v. 5, pp.1-8, 1994 M. Hannan and I. Walker, “Kinematics and the implementation of an elephant’s trunk manipulator and other continuum style robots,” Journal of Robotics Systems, v. 20, n. 2, pp. 45-63, 2003 M. Ciocarlie and P. Allen, “A Design and Analysis Tool for Underactuated Compliant Hands,” Proceedings of the 2009 IEEE/RSJ Int. Conf. on Intelligent Robotics and Systems, pp. 5235-5239, 2009 A. Midha, T. Norton, and L. Howell, “On the nomenclature, classification, and abstractions of compliant mechanisms,” Trans. of the ASME, v. 116, n. 1, pp.270-280, 1994 S. Timoshenko and D. Young, Advanced Dynamics, p. 204, McGrawHill, 1948. A. DeLuca and B. Siciliano, “Closed-form dynamic model of planar multilink lightweight robots,” IEEE Trans. on Systems, Man and Cybernetics, v. 21, n. 4, pp. 826-839, 1991 G. Chirikjian and J. Burdick, “A modal approach to hyper-redundant manipulator kinematics,” IEEE Trans. on Robotics and Automation, v. 10, n. 3, pp. 343-354, 1994 N. Lobontiu, “Distributed-Parameter dynamic model and optimized design of a four-link pendulum with flexure hinges,” Mechanism and Machine Theory, v. 36, pp. 653-669, 2001 N. Lobontiu and E. Garcia, “Two microcantilever designs: lumpedparameter model for static and modal analysis,” Journal of Microelectromechanical Systems, v. 13, 2004 H. Hanafusa and H. Asada, “Stable Prehension by a Robot Hand with Elastic Fingers,” Proceedings of the 7th International Symposium on Industrial Robots, pp. 361-368, 1977. L. Odhner and A. Dollar, “Fast, accurate models for predicting the compliance of elastic flexure-jointed robots,” ASME Internat. Design Engineering Technical Conf and Computers and Information in Engineering Conf., 2010. R. Levien, “The Euler spiral: a mathematical history,” Technical Report No. UCB/EECS-2008-111, University of California at Berkeley, 2008 B. Horn, “The Curve of Least Energy,” ACM Trans. On Mathematical Software, vol. 9, n. 4, pp. 441-460, 1983 W. Press, S. Teukolsky, W. Vetterling, and B. Flannery, Numerical Recipes in C, 2nd ed., p. 190, Cambridge University Press, 1992. R. Levien, “The elastica: a mathematical history,” Technical Report No. UCB/EECS-2008-103, University of California at Berkeley, 2008 L. Odhner and A. Dollar, “The Freeform Manipulator Analysis Tool” [Online]. Avaliable: http://www.eng.yale.edu/grablab/fmat/ S. Crandall, C. Dahl and T. Lardner, An Introduction to the Mechanics 2nd ed., p. 587, McGraw-Hill, 1999. of Solids,

0XOWLSULRULW\ &DUWHVLDQ ,PSHGDQFH &RQWURO 5REHUW 3ODWW -U

0XKDPPDG $EGDOODK &KDUOHV :DPSOHU

&RPSXWHU 6FLHQFH DQG $UWLÀFLDO ,QWHOOLJHQFH /DERUDWRU\ 0DVVDFKXVHWWV ,QVWLWXWH RI 7HFKQRORJ\ USODWW#FVDLOPLWHGX

0DQXIDFWXULQJ 6\VWHPV 5HVHDUFK /DE *HQHUDO 0RWRUV {PXKDPPDGDEGDOODKFKDUOHVZZDPSOHU}#JPFRP

$EVWUDFW³ 0DQLSXODWRU FRPSOLDQFH LV ZHOO NQRZQ WR EH LP SRUWDQW WR URERW PDQLSXODWLRQ DQG DVVHPEO\ 5HFHQWO\ WKLV KDV EHHQ KLJKOLJKWHG E\ WKH GHYHORSPHQW RI QHZ KLJO\FRPSOLDQW URERW PDQLSXODWRUV VXFK DV WKH %DUUHWW DUP RU WKH '/5 OLJKWZHLJKW PDQLSXODWRU > @ ,W LV DOVR FOHDU WKDW GH[WHURXV PDQLSXODWLRQ LQYROYHV WRXFKLQJ WKH HQYLURQPHQW DW GLIIHUHQW ORFDWLRQV VLPXOWDQHRXVO\ SHUKDSV DW GLIIHUHQW SRLQWV RQ WKH URERW KDQG RU ÀQJHUV  ,Q WKHVH VLWXDWLRQV LW LV SDUWLFXODUO\ DWWUDFWLYH WR FRQWURO WKH V\VWHP XVLQJ D PXOWLSULRULW\ VWUDWHJ\ ZKHUH VHYHUDO FRQWDFW SRLQWV DUH FRPPDQGHG LQ SDUDOOHO 0XOWLSULRULW\ &DUWHVLDQ LPSHGDQFH FRQWURO LV WKH QDWXUDO FRPELQDWLRQ RI WKHVH WZR LGHDV 7KH V\VWHP UHDOL]HV VHYHUDO LPSHGDQFHV ZLWK GLIIHUHQW UHIHUHQFH SRVLWLRQV DW GLIIHUHQW SRLQWV RQ WKH URERW ZLWK D VSHFLÀHG RUGHU RI SULRULW\ :H ÀQG D FRQWUROOHU WKDW PLQLPL]HV DQ DUELWUDU\ TXDGUDWLF QRUP RQ WKH VHFRQGSULRULW\ LPSHGDQFH HUURU VXEMHFW WR FRQVWUDLQWV GHULYLQJ IURP WKH ÀUVW SULRULW\ LPSHGDQFH WDVN :H DOVR VKRZ WKDW WKH ORFDOO\ RSWLPDO FRQWUROOHU GRHV QRW UHTXLUH IRUFH IHHGEDFN LQ LWV LPSOHPHQWDWLRQ IRU SDVVLYH GHVLUHG LQHUWLDV 7KH UHVXOWV DUH LOOXVWUDWHG LQ VLPXODWLRQ

, , 1752'8&7,21 6LQFH WKH HDUOLHVW GD\V RI URERWLFV UHVHDUFKHUV KDYH EHHQ DZDUH RI WKH SRVVLELOLW\ RI FRQWUROOLQJ VHFRQGSULRULW\ REMHF WLYHV XVLQJ GHJUHHV RI IUHHGRP '2)V WKDW DUH UHGXQGDQW ZLWK UHVSHFW WR D À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ÀFLW E\ SURSRVLQJ D ORFDOO\RSWLPDO GXDOSULRULW\ LPSHGDQFH ODZ )LUVW ZH SURSRVH D JHQHUDO ODZ WKDW XVHV IRUFH IHHGEDFN WR UHDOL]H DQ DUELWUDU\ GXDOSULRULW\ LPSHGDQFH 7KHQ ZH FKDUDFWHUL]H WKH VXEVHW RI LPSHGDQFHV WKDW FDQ EH UHDOL]HG XVLQJ WRUTXH FRQWUROOHG PDQLSXODWRUV ZLWKRXW IRUFH VHQVRUV ,Q FRQWUDVW WR SULRU ZRUN ZH FRQVLGHU WKH VSDFH RI DOO SRVVLEOH FRQWURO ODZV WKDW UHDOL]H WKH ÀUVWSULRULW\ LPSHGDQFH DQG VHOHFW WKH RQH WKDW LV RSWLPDO ZLWK UHVSHFW WR DQ DUELWUDU\ TXDGUDWLF RSWLPL]DWLRQ FULWHULRQ 7KH PXOWLSULRULW\ DSSURDFK WR URERW PDQLSXODWRU FRQWURO KDV EHHQ VWXGLHG H[WHQVLYHO\ 0DQ\ HDUO\ DSSURDFKHV FDOFXODWHG  3RUWLRQV

MRLQW YHORFLWLHV WKDW DWWHPSW WR DFKLHYH D VHFRQGSULRULW\ REMHF WLYH VXFK DV DYRLGLQJ REVWDFOHV RU PDQLSXODWRU VLQJXODULWLHV ZKLOH DFKLHYLQJ D GHVLUHG HQGHIIHFWRU YHORFLW\ >  @ &KLDYHULQL FRQVLGHUV D GDPSHGOHDVWVTXDUHV YHUVLRQ RI WKH FRQWURO ODZ WKDW LV UREXVW WR DOJRULWKPLF VLQJXODULWLHV >@ $QWRQHOOL SURYLGHV D /\DSXQRY DQDO\VLV GHPRQVWUDWLQJ WKDW WKH EDVLF DSSURDFK LV VWDEOH >@ 5HODWHG DSSURDFKHV KDYH EHHQ DSSOLHG LQ PRUH JHQHUDO FRQWH[WV > @ $ VLJQLÀFDQW ERG\ RI ZRUN H[SORUHV WKH JHQHUDO SUREOHP RI LPSHGDQFH FRQWURO LQ WKH FRQWH[W RI UHGXQGDQW PDQLSXODWRUV %XLOGLQJ RQ +RJDQ·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ÀUVW SULRULW\ WDVN LQGHSHQGHQW RI WKH ORZHUSULRULW\ WDVNV LW GRHV QRW RSWLPL]H IRU VHFRQGSULRULW\ SHUIRUPDQFH 7KH FXUUHQW SDSHU H[SORUHV WKLV SRLQW :H SURYLGH D FRQWURO ODZ WKDW PLQLPL]HV TXDGUDWLF HUURU ZLWK UHVSHFW WR WKH GHVLUHG ORZHUSULRULW\ LPSHGDQFH ZKLOH VWLOO HQVXULQJ LQGHSHQGHQFH RI WKH ÀUVWSULRULW\ REMHFWLYH ,W WXUQV RXW WKDW IRU WKH RSWLPDO GXDOSULRULW\ FRQWUROOHU LW LV SRVVLEOH WR LGHQWLI\ WKH RSHUDWLRQDO VSDFH GLUHFWLRQV ZKHUH WKH VHFRQGSULRULW\ WDVN LV LQGHSHQGHQW RI WKH ÀUVWSULRULW\ WDVN 7KHVH LQGHSHQGHQW GLUHFWLRQV FDQ EH DGMXVWHG E\ FKDQJLQJ WKH ZHLJKWLQJ PDWUL[ XVHG E\ WKH RSWLPL]DWLRQ FULWHULRQ ,, %$&.*5281' $ *HQHUDOL]HG HTXDWLRQ RI PRWLRQ 7KH G\QDPLF PRWLRQ RI D URERW DUP ZLWK n UHYROXWH MRLQWV LV W\SLFDOO\ XQGHUVWRRG LQ WHUPV RI WKH IROORZLQJ JHQHUDOL]HG

RI WKLV SXEOLFDWLRQ KDYH SDWHQWV SHQGLQJ

 145

HTXDWLRQ RI PRWLRQ M q¨ + η = τa + τ, ZKHUH M LV WKH n × n PDQLSXODWRU LQHUWLD PDWUL[ q¨ LV DQ nYHFWRU RI PDQLSXODWRU MRLQW DFFHOHUDWLRQV τ LV DQ nYHFWRU RI MRLQW WRUTXHV UHVXOWLQJ IURP H[WHUQDOO\ DSSOLHG ORDGV τa LV D YHFWRU RI DFWXDWRU WRUTXHV DQG η GHVFULEHV WKH VXP RI IULFWLRQDO FRULROLV FHQWULIXJDO DQG JUDYLWDWLRQDO WRUTXHV >@ 7KH GHSHQGHQFH RI WKHVH WHUPV RQ PDQLSXODWRU FRQÀJXUDWLRQ LV LPSOLFLWO\ DVVXPHG ,Q RUGHU WR VLPSOLI\ QRWDWLRQ ZH LQWURGXFH WKH IROORZLQJ VXEVWLWXWLRQ

DV FRPSOHWHO\ DV SRVVLEOH JLYHQ WKH PHFKDQLFV RI WKH PDQLS XODWRU ZKHUH ˜1 ,  f1∗ = B1 x˙ 1 + K1 x Ω1 LV WKH GHVLUHG LQHUWLD B1 LV WKH GHVLUHG GDPSLQJ K1 LV WKH GHVLUHG VWLIIQHVV DQG x ˜1 LV WKH SRVH HUURU LQ WKH ÀUVWSULRULW\ 325 ,Q DGGLWLRQ GXDOSULRULW\ &DUWHVLDQ LPSHGDQFH FRQWURO PXVW DOVR PLQLPL]H HUURU ZLWK UHVSHFW WR WKH VHFRQGSULRULW\ LPSHGDQFH REMHFWLYH Ω2 x ¨2 + f2∗ = f2 ,



τa = u + η,

WR WKH JUHDWHVW H[WHQW SRVVLEOH ZKLOH QRW LPSDFWLQJ SHUIRU PDQFH ZLWK UHJDUG WR WKH ÀUVWSULRULW\ REMHFWLYH ZKHUH

VXFK WKDW WKH HTXDWLRQ RI PRWLRQ FDQ EH H[SUHVVHG LQ WHUPV RI D FRPPDQG YHFWRU u

˜2 , f2∗ = B2 x˙ 2 + K2 x

M q¨ = u + τ.

Ω2 LV WKH GHVLUHG LQHUWLD B2 LV WKH GHVLUHG GDPSLQJ K2 LV WKH GHVLUHG VWLIIQHVV DQG x ˜2 LV WKH SRVH HUURU DW WKH VHFRQGSULRULW\ 325



% 2SHUDWLRQDO VSDFH ,W LV IUHTXHQWO\ XVHIXO WR GHVLJQ FRQWUROOHUV GHÀQHG LQ RSHUDWLRQDO VSDFH FRRUGLQDWHV UDWKHU WKDQ LQ MRLQW VSDFH :KLOH WKH WHUP ´RSHUDWLRQDO VSDFHµ PD\ UHIHU WR DQ\ FRRUGLQDWH V\VWHP UHOHYDQW WR D URERW WDVN LW XVXDOO\ UHIHUV WR WKH VSDFH RI SRVLWLRQV DQG RULHQWDWLRQV RI WKH HQGHIIHFWRU UHSUHVHQWHG E\ D SDUDPHWHUL]DWLRQ RI SE(3) ,Q WKLV SDSHU SE(3) LV SDUDPHWHU L]HG XVLQJ H[SRQHQWLDO FRRUGLQDWHV ZKHUHE\ D &DUWHVLDQ SRVH LV HQFRGHG E\ D YHFWRU ZLWK WKH ÀUVW WKUHH QXPEHUV GHVFULELQJ SRVLWLRQ DQG WKH ODVW WKUHH QXPEHUV GHVFULELQJ RULHQWDWLRQ XVLQJ WKH D[LVDQJOH UHSUHVHQWDWLRQ >@ 7KH &DUWHVLDQ YH ORFLW\ RI WKH HQGHIIHFWRU ZLOO EH UHSUHVHQWHG DV D WZLVW DQG WKH DFFHOHUDWLRQ DV WKH GHULYDWLYH RI WZLVW 6LPLODUO\ ORDGV LQ &DUWHVLDQ VSDFH ZLOO EH ZULWWHQ DV ZUHQFKHV VL[YHFWRU WKDW FRQFDWHQDWHV D IRUFH DQG D PRPHQW >@ 7KH HQGHIIHFWRU -DFRELDQ J UHODWHV MRLQW YHORFLWLHV q ˙ WR &DUWHVLDQ WZLVWV DW WKH HQGHIIHFWRU x ˙ x˙ = J q ˙ ,,, ' 8$/ 35,25,7< & $57(6,$1 63$&( ,03('$1&( &21752/

'XDO SULRULW\ &DUWHVLDQ LPSHGDQFH FRQWURO LV GHÀQHG ZLWK UHVSHFW WR WZR SRLQWV RI UHIHUHQFH 325V  $ 325 LV D UHIHUHQFH IUDPH DWWDFKHG WR WKH PDQLSXODWRU WKDW LV WKH RE MHFW RI &DUWHVLDQ FRQWURO ,Q JHQHUDO GXDO SULRULW\ &DUWHVLDQ LPSHGDQFH FRQWURO UHDOL]HV WZR GLIIHUHQW LPSHGDQFH ODZV GH ÀQHG ZLWK UHVSHFW WR WZR GLIIHUHQW 325V VLPXOWDQHRXVO\ 2QH LPSHGDQFH REMHFWLYH LV ÀUVWSULRULW\ ZKLOH WKH RWKHU LV VHFRQG SULRULW\ 'XDOSULRULW\ &DUWHVLDQ LPSHGDQFH FRQWURO PLQLPL]HV HUURU ZLWK UHVSHFW WR WKH ÀUVWSULRULW\ LPSHGDQFH REMHFWLYH ZKLOH DOVR UHDOL]LQJ WKH VHFRQGSULRULW\ REMHFWLYH WR WKH JUHDW HVW H[WHQW SRVVLEOH $ 3UREOHP VWDWHPHQW ¨1  DQG f1 EH WKH SRVH WZLVW DFFHOHUDWLRQ /HW x1  x˙ 1  x DQG H[WHUQDOO\ DSSOLHG ZUHQFK DVVRFLDWHG ZLWK WKH ÀUVWSULRULW\ ¨2  DQG f2 VLPLODUO\ IRU WKH VHFRQG 325 'HÀQH x2  x˙ 2  x SULRULW\ 325 7KH JRDO RI GXDOSULRULW\ LPSHGDQFH FRQWURO LV WR UHDOL]H WKH ÀUVWSULRULW\ LPSHGDQFH REMHFWLYH Ω1 x ¨1 + f1∗ = f1 ,





146



% 2SWLPL]DWLRQ FULWHULRQ 7KH SRVVLELOLW\ RI UHDOL]LQJ WZR &DUWHVLDQ VSDFH LPSHGDQFHV GHSHQGV RQ WKH GHJUHH RI PRELOLW\ EHWZHHQ WKH WZR 325V ,Q JHQHUDO LW LV QRW SRVVLEOH WR UHDOL]H ERWK &DUWHVLDQ LPSHGDQFHV FRQFXUUHQWO\ EHFDXVH WKH WZR 325V DUH QRW QHFHVVDULO\ VXI ÀFLHQWO\ PRELOH ZLWK UHVSHFW WR HDFK RWKHU 7KHUHIRUH WKH PXOWLSOHSULRULW\ DSSURDFK DSSOLHV DFWXDWRU WRUTXHV VXFK WKDW WKH SULPDU\ LPSHGDQFH REMHFWLYH LV DFKLHYHG ZKLOH PLQLPL]LQJ DQ RSWLPL]DWLRQ FULWHULRQ DVVRFLDWHG ZLWK WKH VHFRQGSULRULW\ LPSHGDQFH REMHFWLYH :H UHVWULFW RXU DWWHQWLRQ WR WKH FDVH ZKHUH WKH RSWLPL]DWLRQ FULWHULRQ LV WKH ZHLJKWHG VTXDUHG PDJQLWXGH RI WKH &DUWHVLDQ DFFHOHUDWLRQ DW WKH VHFRQGSULRULW\ 325 (¨ x2 − x ¨′2 )T W (¨ x2 − x ¨′2 ) T = z z,

ǫ =



ZKHUH x ¨′2 LV WKH GHVLUHG &DUWHVLDQ VSDFH DFFHOHUDWLRQ DW WKH VHFRQGSULRULW\ 325 DQG 1

x2 − x ¨′2 ). z = W 2 (¨



7KH JRDO RI GXDOSULRULW\ &DUWHVLDQ LPSHGDQFH FRQWURO LV WR PLQLPL]H (TXDWLRQ  ZKLOH UHDOL]LQJ WKH ÀUVWSULRULW\ LPSHGDQFH LQ (TXDWLRQ  & &RQWURO ODZ 7KH ÀUVWSULRULW\ &DUWHVLDQ LPSHGDQFH FRQVWUDLQV WKH VSDFH RI MRLQW DFFHOHUDWLRQV WKDW PXVW EH FRQVLGHUHG 6XEVWLWXWLQJ ¨1 LQ (TXDWLRQ  DQG WDNLQJ WKH SVHXGRLQYHUVH J1 q¨ + J˙1 q˙ IRU x ZH KDYH / 0 ∗ ˙ q¨ = J1+ Ω−1  1 (f1 − f1 ) − J q˙ + N1 λ, ¨2 = J2 q¨ + ZKHUH N1 = I − J1+ J1 DQG λ LV DUELWUDU\ 8VLQJ x J˙2 q ˙ WKH VSDFH RI &DUWHVLDQ DFFHOHUDWLRQV DW WKH VHFRQGSULRULW\ 325 FRQVLVWHQW ZLWK WKH ÀUVWSULRULW\ LPSHGDQFH LV 0 / ∗ ˙ (f − f ) − J q ˙ + J˙2 q˙ + J2 N1 λ.  x ¨2 = J2 J1+ Ω−1 1 1 1

$PRQJ WKRVH MRLQW DFFHOHUDWLRQV SHUPLWWHG E\ WKH FRQVWUDLQW ZH PXVW ÀQG RQH WKDW PLQLPL]HV WKH RSWLPL]DWLRQ FULWHULRQ 7KLV FDQ EH DFKLHYHG E\ VROYLQJ (TXDWLRQ  IRU WKH YDOXH RI λ WKDW PLQLPL]HV ǫ (TXDWLRQ   RU HTXLYDOHQWO\ z (TXDWLRQ   6ROYLQJ (TXDWLRQ  IRU WKH GHVLUHG DFFHOHUDWLRQ DW WKH VHFRQG 325 ZH KDYH x ¨′2

=

Ω−1 2 (f2

−

f2∗ ).

W (¨ x −x ¨′2 )   /2 1 ∗ ˙ (f − f ) − J q ˙ + J˙2 q˙ = W 2 J2 J1+ Ω−1 1 1 1 1  ∗  −Ω−1 2 (f2 − f2 ) + J2 N1 λ .

7KH YDOXH RI λ WKDW PLQLPL]HV z T z FDQ EH IRXQG XVLQJ WKH SVHXGRLQYHUVH 0 / ∗ ˙2 q˙ (f − f ) − J λ = Jˆ2 Ω−1 2 2 2 / 0 −1 + ˆ −J2 J2 J1 Ω1 (f1 − f1∗ ) − J˙1 q˙ +(I − Jˆ2 J2 N1 )β,



ZKHUH 1 1 Jˆ2 = (W 2 J2 N1 )+ W 2 .



7KH WHUP (I − Jˆ2 J2 N1 )β IRU DUELWUDU\ β UHSUHVHQWV WKH DGGLWLRQDO IUHHGRPV OHIW RYHU LQ WKH FDVH WKDW WKH VHFRQG SULRULW\ RSWLPL]DWLRQ FULWHULRQ GRHV QRW FRPSOHWHO\ GHWHUPLQH WKH PRWLRQ RI DOO MRLQWV 7KH PDWUL[ (I − Jˆ2 J2 N1 ) LV DQ RUWKRJRQDO SURMHFWLRQ PDWUL[ WKDW VSDQV WKH MRLQW DFFHOHUDWLRQV LUUHOHYDQW WR HLWKHU REMHFWLYH 7KH MRLQW DFFHOHUDWLRQV WKDW UHDOL]H WKH ÀUVWSULRULW\ LPSHGDQFH (TXDWLRQ  ZKLOH DOVR PLQLPL]LQJ WKH RSWLPL]D WLRQ FULWHULRQ ZLWK UHVSHFW WR WKH VHFRQGSULRULW\ LPSHGDQFH (TXDWLRQ  FDQ EH IRXQG E\ VXEVWLWXWLQJ λ EDFN LQWR (TXD WLRQ  / 0 ∗ ˙ (f − f ) − J q ˙ q¨ = J1+ Ω−1 1 1 1 1 0 / −1 ˆ +N1 J2 Ω2 (f2 − f2∗ ) − J˙2 q˙ / 0 ∗ ˙ −N1 Jˆ2 J2 J1+ Ω−1 1 (f1 − f1 ) − J1 q˙ +N1 (I − Jˆ2 J2 N1 )β.





ˆ2 = (I − Jˆ2 J2 ). N

x ¨′2

1 2

=

ˆ2 N1 β. +N1 N

ZKHUH

6XEVWLWXWLQJ (TXDWLRQ  LQWR (TXDWLRQ  IRU x ¨2 DQG XVLQJ DERYH ZH KDYH z

$V D UHVXOW (TXDWLRQ  EHFRPHV / 0 ˆ2 J + Ω−1 (f1 − f ∗ ) − J˙1 q˙ q¨ = N 1 1 1 0 / ∗ ˙ +Jˆ2 Ω−1 2 (f2 − f2 ) − J2 q˙



6XEVWLWXWLQJ LQWR WKH HTXDWLRQ RI PRWLRQ (TXDWLRQ  IRU q¨ ZH KDYH u

M q¨ − τ 0 / ˆ2 J + Ω−1 (f1 − f ∗ ) − J˙1 q˙ = MN 1 1 1 0 / −1 ∗ ˆ +M J2 Ω2 (f2 − f2 ) − J˙2 q˙

=

ˆ2 N1 β − τ. +M N1 N



,Q RUGHU WR UHDOL]H (TXDWLRQ  LQ SUDFWLFH LW LV QHFHVVDU\ WR PHDVXUH WKH H[WHUQDOO\ DSSOLHG ORDGV 7KLV PLJKW EH DF FRPSOLVKHG E\ PHDVXULQJ WKH ORDGV f1 DQG f2  DSSOLHG DW WKH ÀUVW DQG VHFRQGSULRULW\ 325V DQG DVVXPLQJ WKDW QR H[WHUQDO ORDGV DUH DSSOLHG HOVHZKHUH RQ WKH PDQLSXODWRU $FFRUGLQJO\ ZH KDYH τ = J1T f1 + J2T f2 . 6XEVWLWXWLQJ LQWR (TXDWLRQ  ZH KDYH u

= M q¨ − τ 0 / ˆ2 J + Ω−1 (f1 − f ∗ ) − J˙1 q˙ = MN 1 1 1 / 0 ∗ ˙ +M Jˆ2 Ω−1 2 (f2 − f2 ) − J2 q˙

ˆ2 N1 β − J T f1 − J T f2 . +M N1 N 2 1



7KH FRQWURO ODZ LQ (TXDWLRQ  VROYHV WKH SUREOHP SURSRVHG LQ 6HFWLRQ ,,,$ ZKLOH PLQLPL]LQJ WKH &DUWHVLDQ DFFHOHUDWLRQ RI WKH VHFRQGSULRULW\ 325 DQG DVVXPLQJ WKDW H[WHUQDOO\ DSSOLHG ZUHQFKHV DW HDFK 325 FDQ EH PHDVXUHG GLUHFWO\ ' $QDO\VLV 7KH FORVHGORRS EHKDYLRU RI WKH GXDOSULRULW\ &DUWHVLDQ LPSHGDQFH FRQWUROOHU (TXDWLRQ  FDQ EH XQGHUVWRRG LQ WHUPV RI WKH &DUWHVLDQ DFFHOHUDWLRQV DW WKH ÀUVW DQG VHFRQGSULRULW\ 325V )LUVW QRWH WKDW J1 Jˆ2

1

1

= J1 (W 2 J2 N1 )+ W 2 1

1

1

1

= J1 N1 J2T W 2 (W 2 J2 N1 J2T W 2 )+ W 2 = 0.

7KLV HTXDWLRQ FDQ EH VLPSOLÀHG E\ XVLQJ WKH IDFW WKDW IRU DQ\ PDWUL[ A WKH IROORZLQJ LV WUXH A+ = AT (AAT )+  7KHUHIRUH ZH KDYH WKDW N1 Jˆ2

=

1 2

N1 (W J2 N1 )+ W 1 2

1 2 1 2

1 2

=

N1 (N1 J2T W )(W J2 N1 J2T W )+ W

=

(N1 J2T W

=

Jˆ2 .

1 2

)(W

1 2

J2 N1 J2T W

1 2

+

) W

1 2

7KHUHIRUH ZH KDYH WKDW ˆ2 J1 N

= J1 (I − Jˆ2 J2 ) = J1 .



8VLQJ WKH DERYH UHVXOWV DQG VXEVWLWXWLQJ (TXDWLRQ  LQWR x ¨1 = J1 q¨ + J˙1 q ˙ WKH &DUWHVLDQ DFFHOHUDWLRQ DW WKH ÀUVW 325 LV

1 2



 147

∗ x ¨1 = Ω−1 1 (f1 − f1 ),

WKHUHE\ UHDOL]LQJ WKH GHVLUHG ÀUVWSULRULW\ LPSHGDQFH 7KH &DUWHVLDQ DFFHOHUDWLRQ RI WKH VHFRQGSULRULW\ 325 FDQ EH FDOFXODWHG XVLQJ x ¨2 = J2 q¨ + J˙2 q ˙ / 0 ∗ ˙ ˙ x ¨2 = J2 Jˆ2 Ω−1 2 (f2 − f2 ) − J2 q˙ + J2 q˙ 0 / ∗ ˙ +(I − J2 Jˆ2 )J2 J1+ Ω−1 1 (f1 − f1 ) − J1 q˙

6XEVWLWXWLQJ EDFN LQWR WKH HTXDWLRQ RI PRWLRQ (TXDWLRQ  DQG VROYLQJ IRU x ¨1  ZH DUULYH DW (TXDWLRQ  GHPRQVWUDWLQJ WKDW WKH VHFRQGSULRULW\ REMHFWLYH GRHV QRW LQÁXHQFH WKH G\QDPLFV RI WKH ÀUVWSULRULW\ REMHFWLYH +RZHYHU VXEVWLWXWLQJ LQWR WKH HTXDWLRQ RI PRWLRQ DQG VROYLQJ IRU x ¨2  ZH JHW

7KH ODVW WHUP RI WKLV HTXDWLRQ FDQ EH HOLPLQDWHG E\ QRWLQJ

,W VKRXOG EH QRWHG WKDW WKH VHFRQG WHUP DERYH LV QRW RUWKRJRQDO WR WKH ÀUVW WHUP DV LV WKH FDVH LQ (TXDWLRQ  7KLV EHJV WKH TXHVWLRQ RI ZKHWKHU WKLV FRQWURO ODZ LV RSWLPDO IRU DQ\ FULWHULRQ RI WKH IRUP RI (TXDWLRQ  7KLV TXHVWLRQ FDQ EH DQVZHUHG XVLQJ WKH /DJUDQJH PXOWLSOLHU PHWKRG :H DUH LQWHUHVWHG LQ FRQWURO ODZV WKDW PLQLPL]H

x ¨2

=

ˆ2 N1 β. +J2 N1 N

ˆ2 N1 J 2 N1 N

= = = =

=

J2 N1 (I − Jˆ2 J2 )N1 J2 N1 − J2 N1 Jˆ2 J2 N1 1

1

J2 N1 − J2 N1 (W 2 J2 N1 )+ W 2 J2 N1 / 1 1 W − 2 W 2 J2 N1 0 1 1 1 −W 2 J2 N1 (W 2 J2 N1 )+ W 2 J2 N1

f (¨ q ) = (J2 q¨ − β)T W (J2 q¨ − β) VXEMHFW WR

0.

7KHUHIRUH WKH UHVXOWLQJ FORVHGORRS LPSHGDQFH DW x2 LV / 0 ∗ ˙ ˙ x ¨2 = J2 Jˆ2 Ω−1 2 (f2 − f2 ) − J2 q˙ + J2 q˙ 0 / ∗ ˙ +(I − J2 Jˆ2 )J2 J1+ Ω−1 1 (f1 − f1 ) − J1 q˙ . 

(TXDWLRQ  FDQ EH XQGHUVWRRG E\ UHFRJQL]LQJ WKDW J2 Jˆ2 DQG (I − J2 Jˆ2 ) DUH RUWKRJRQDO ZHLJKWHG SURMHFWLRQ PDWULFHV J2 Jˆ2 (I − J2 Jˆ2 ) = 0 7KH ÀUVW WHUP LV SURMHFWHG WKURXJK J2 Jˆ2 DQG GHVFULEHV WKH FRQWULEXWLRQ RI WKH VHFRQGSULRULW\ LPSHGDQFH WR WKH FORVHGORRS DFFHOHUDWLRQ RI x2  7KH ODVW WHUP LV SURMHFWHG WKURXJK (I − J2 Jˆ2 ) DQG GHVFULEHV WKH HIIHFW WKDW WKH ÀUVWSULRULW\ LPSHGDQFH REMHFWLYH KDV RQ WKH FORVHG ORRS EHKDYLRU DW WKH VHFRQGSULRULW\ 325 1RWLFH WKDW WKH VHFRQG SULRULW\ LPSHGDQFH RSHUDWHV FRPSOHWHO\ LQGHSHQGHQWO\ LQ WKH UDQJH VSDFH RI J2 Jˆ2  0XOWLSO\LQJ ERWK VLGHV RI (TXDWLRQ  E\ J2 Jˆ2  ZH KDYH ∗ ¨2 = J2 Jˆ2 Ω−1 J2 Jˆ2 x 2 (f2 − f2 ).

∗ J2 M −1 J1T Λ1 Ω−1 1 (f1 − f1 ) −1 T +J2 M −1 J2|1 Λ2|1 Ω2 (f2 − f2∗ ).

g(¨ q ) = J1 q¨ − α = 0, ∗ ˙ Ω−1 1 (f1 − f1 ) − J1 q˙

∗ ˙ ˙ ZKHUH α = DQG β = Ω−1 2 (f2 − f2 ) − J2 q 'LIIHUHQWLDWLQJ f DQG g DQG FRQVWUXFWLQJ WKH /DJUDQJLDQ ZH KDYH  hT J1 + q¨T J2T W J2 − β T W J2 = 0,

ZKHUH h LV WKH /DJUDQJH PXOWLSOLHU (TXDWLRQ  PXVW EH VDWLVÀHG IRU VRPH YDOXH RI h DW DQ RSWLPXP (TXLYDOHQWO\ ZH FDQ UHTXLUH   N1 J2T W J2 q¨ − J2T W β = 0.

6ROYLQJ IRU q¨ E\ VXEVWLWXWLQJ (TXDWLRQ  LQWR (TXDWLRQ  DQG VXEVWLWXWLQJ LQWR WKH DERYH WKH FRQGLWLRQ EHFRPHV 0   / T Λ2|1 β − J2T W β = 0. N1 J2T W J2 M −1 J1T Λ1 α + J2|1

6LQFH ZH PXVW DVVXPH α WR EH DUELWUDU\ WKH FRQGLWLRQ LV RQO\ VDWLVÀHG ZKHQ



N1 J2T W J2 M −1 J1T Λ1 α = 0.

( &RPSDULVRQ WR 6HQWLV.KDWLE 7KHVH FORVHGORRS G\QDPLFV PD\ EH FRPSDUHG WR WKH G\ QDPLFV RI WKH PXOWLSULRULW\ 6HQWLV.KDWLE FRQWURO ODZ %DVHG RQ >@ WKH 6HQWLV.KDWLE FRQWURO ODZ LV / 0 ∗ ˙ τ = J1T Λ1 Ω−1 1 (f1 − f1 ) − J1 q˙ 0 / ∗ T ˙ +J2|1  Λ2|1 Ω−1 2 (f2 − f2 ) − J2 q˙ . ,Q WKH DERYH FRQWURO ODZ ZH KDYH XVHG

T −1 Λ2|1 = (J2|1 M −1 J2|1 )



+RZHYHU VLQFH WKHUH LV QR À[HG W VXFK WKDW (TXDWLRQ  LV WUXH IRU DUELWUDU\ α (TXDWLRQ  FDQQRW EH RSWLPDO ,9 , 03('$1&( &21752/ :,7+287 )25&( 6(16256 $Q LPSRUWDQW FRQFHUQ ZLWK PXOWLSULRULW\ LPSHGDQFH FRQWURO DV LW LV H[SUHVVHG LQ (TXDWLRQ  LV WKH QHHG WR VHQVH H[WHUQDOO\ DSSOLHG ORDGV DW WZR GLIIHUHQW SRLQWV RQ WKH PDQLSXODWRU :KLOH WKLV PD\ EH IHDVLEOH LQ VRPH FDVHV LW LV LPSRUWDQW WR DGGUHVV WKH FDVH ZKHUH VXFK IHHGEDFN LV QRW DYDLODEOH $ &RQWURO ODZ

DQG J2|1 = J2 NM , ZKHUH NA LV WKH LQHUWLDZHLJKWHG QXOO VSDFH RI WKH ÀUVW SULRULW\ REMHFWLYH NM = I − M −1 J1T (J1 M −1 J1T )−1 J1 .

 148

&RQVLGHU (TXDWLRQ  DQG VXSSRVH WKDW WKH MRLQW WRUTXH FDXVHG E\ DOO H[WHUQDOO\ DSSOLHG ORDGV τ LV VXEVWLWXWHG IRU HDFK LQVWDQFH RI J1T f1 RU J2T f2  /HW Λ1 = (J1 M −1 J1T )−1 EH WKH SDVVLYH PDQLSXODWRU LQHUWLD GHVFULEHG LQ &DUWHVLDQ FRRUGLQDWHV DW WKH ÀUVWSULRULW\ 325 /HW Λ2 = (J2 M −1 J2T )−1 EH WKH VLPLODU TXDQWLW\ DW WKH VHFRQGSULRULW\ 325 6HW WKH

GHVLUHG &DUWHVLDQ LQHUWLDV WR WKHVH SDVVLYH YDOXHV Ω1 = Λ1 DQG Ω2 = Λ2  7KHQ (TXDWLRQ  EHFRPHV 0 / ˆ2 J + Λ−1 f ∗ + J˙1 q˙ u = −M N 1 1 1 0 / ∗ ˙ −M Jˆ2 Λ−1 2 f2 + J2 q˙ ˆ2 J + J1 M −1 τ +M N 1 +M Jˆ2 J2 M −1 τ ˆ2 N1 β +M N1 N

−τ.

7KLV H[SUHVVLRQ FDQ EH VLPSOLÀHG XVLQJ (TXDWLRQ  WR ÀQG WKDW

=

x ¨2

ˆ2 J + (Λ−1 f ∗ + J˙1 q) J2 M −1 τ − J2 N ˙ 1 1 1 −1 ∗ −J2 Jˆ2 (Λ f + J˙2 q) ˙ + J˙2 q˙

=

1

2

∗ ˙ ˙ J2 M −1 τ − (I − J2 Jˆ2 )J1+ (Λ−1 1 f1 + J1 q) −1 ∗ −J2 Jˆ2 Λ f2 + (I − J2 Jˆ2 )J˙2 q. ˙

=

∗ J2 Jˆ2 x ¨2 = J2 Jˆ2 (J2 M −1 τ − Λ−1 2 f2 ).

$V D UHVXOW (TXDWLRQ  FDQ EH UHZULWWHQ ZLWKRXW WKH τ WHUPV 0 / ˆ2 J + Λ−1 f1∗ + J˙1 q˙ u = −M N 1 1 / 0 ∗ ˙ −M Jˆ2 Λ−1 2 f2 + J2 q˙ 

7KLV FRQWURO ODZ KDV WKH VDPH FORVHGORRS LPSHGDQFH DV GRHV (TXDWLRQ  LQ WKH VSHFLDO FDVH WKDW Ω1 = Λ1 DQG Ω2 = Λ2  6LQFH LW GRHV QRW UHTXLUH &DUWHVLDQ IRUFH VHQVRUV WR EH PRXQWHG DW WKH ÀUVW DQG VHFRQGSULRULW\ 325V WKLV FRQWURO ODZ ZLOO EH UHIHUUHG WR DV WKH ´]HURIRUFHµ GXDOSULRULW\ FRQWURO ODZ 7KH IRUFHV DUH QRW ]HUR EXW ZH GR QRW QHHG WR PHDVXUH WKHP % $QDO\VLV 7KH DQDO\VLV RI WKH ]HURIRUFH GXDOSULRULW\ &DUWHVLDQ LPSHGDQFH FRQWURO ODZ (TXDWLRQ  LV VLPLODU WR WKH DQDO\VLV SHUIRUPHG LQ 6HFWLRQ ,,,' 6XEVWLWXWLQJ (TXDWLRQ  LQWR WKH HTXDWLRQ RI PRWLRQ (TXDWLRQ  DQG XVLQJ WKH IDFW WKDW ˙ WKH G\QDPLFV RI WKH ÀUVWSULRULW\ 325 DUH x ¨1 = J1 q¨ + J˙1 q ¨1 + f1∗ = Λ1 J1 M −1 τ. Λ1 x

6XEVWLWXWLQJ (TXDWLRQ  LQWR WKH HTXDWLRQ RI PRWLRQ (TXD WLRQ  DQG XVLQJ WKH IDFW WKDW x ¨2 = J2 q¨ + J˙2 q ˙ WKH G\QDPLFV RI WKH ÀUVWSULRULW\ 325 DUH

0XOWLSO\LQJ WKH DERYH E\ J2 Jˆ2  WKH UHVXOWLQJ FORVHGORRS EHKDYLRU DW WKH VHFRQGSULRULW\ 325 LV

ˆ2 N1 γ M M −1 τ − τ + N ˆ N2 N1 γ.

ˆ2 N1 γ. +M N

Λ1 J1 M −1 τ = f1 + Λ1 J1 M −1 J2T f2 .

1

N1 (I − Jˆ2 J2 )N1 (N1 − N1 Jˆ2 J2 )N1 = (N1 − Jˆ2 J2 )N1 = (I − Jˆ2 J2 )N1 ˆ2 N1 . = N

= =

7KHUHIRUH WKH ODVW IRXU WHUPV RI (TXDWLRQ  FDQ EH VLPSOLÀHG DV IROORZV 0 / ˆ2 N1 γ ˆ2 N1 M −1 τ − τ + N ˆ2 J + J1 + Jˆ2 J2 + N M N 1 / 0 ˆ2 + Jˆ2 J2 M −1 τ − τ + N ˆ2 N1 γ = M N =

DQG (TXDWLRQ  LV LGHQWLFDO WR (TXDWLRQ  :KHQ H[WHUQDO ORDGV DUH DSSOLHG WR ERWK 325V VXFK WKDW τ = J1T f1 + J2T f2  WKHQ ORDGV DSSOLHG WR WKH VHFRQGSULRULW\ 325 f2  DUH SUR MHFWHG WKURXJK Λ1 J1 M −1 J2T RQWR WKH ÀUVWSULRULW\ 325



6LQFH WKH YDOXH RI β GRHV QRW DIIHFW WKH FORVHGORRS EHKDYLRU RI WKH FRQWURO ODZ ZLWK UHVSHFW WR WKH ÀUVW RU VHFRQGSULRULW\ LPSHGDQFH REMHFWLYHV VHW β = M −1 τ +γ ZKHUH γ LV DUELWUDU\ 7KHQ WKH ODVW IRXU WHUPV RI (TXDWLRQ  DUH 0 / ˆ2 N1 γ. ˆ2 N1 M −1 τ − τ + N1 N ˆ2 J + J1 + Jˆ2 J2 + N1 N M N 1 ˆ 2 N1 N1 N

1RWH WKH VLPLODULWLHV EHWZHHQ WKH FORVHGORRS LPSHGDQFH DERYH DQG WKH GHVLUHG LPSHGDQFH LQ (TXDWLRQ  :KHQ DOO H[WHUQDO ORDGV DUH DSSOLHG WR WKH ÀUVWSULRULW\ 325 VXFK WKDW τ = J1T f1  WKHQ Λ1 J1 M −1 τ = f1 .



 149



,I DOO H[WHUQDO ORDGV DUH DSSOLHG WR WKH VHFRQGSULRULW\ 325 WKHQ τ = J2T f2 DQG (TXDWLRQ  EHFRPHV (TXDWLRQ  IRU Ω2 = Λ2  ,I ORDGV DUH DSSOLHG WR ERWK 325V τ = J1T f2 +J2T f2  WKHQ WKH ORDG DSSOLHG WR WKH ÀUVWSULRULW\ 325 LV SURMHFWHG RQWR WKH VHFRQGSULRULW\ 325 9 6 ,08/$7,21 7KH GXDOSULRULW\ DSSURDFK SURSRVHG LQ WKLV SDSHU LV FRP SDUHG WR ERWK D VLQJOHSULRULW\ DSSURDFK DQG WKH 6HQWLV.KDWLE GXDOSULRULW\ FRQWURO ODZ 6HFWLRQ ,,,( LQ WKH FRQWH[W RI D VLPSOH UHJUDVS SUREOHP $ 2YHUYLHZ )LJXUH  LOOXVWUDWHV WKH VLPXODWLRQ VFHQDULR %RWK ÀQJHUV DUH LQWLDOO\ LQ FRQWDFW ZLWK DQ REMHFW DQG WRXFKLQJ LW OLJKWO\ WKH HQWLUH PDQLSXODWRU LV LOOXVWUDWHG LQ )LJXUH   7KH FRQWURO SUREOHP LV WR PRYH RQH ÀQJHU ZKLOH WKH RWKHU ÀQJHU DSSOLHV D VPDOO LQZDUG IRUFH UHVXOWLQJ IURP D FRQVWDQW LPSHGDQFH )RU H[DPSOH WKLV EHKDYLRU FRXOG EH LPSRUWDQW LQ DQ DVVHPEO\ DSSOLFDWLRQ ZKHUH RQH ÀQJHU QHHGV WR PDLQWDLQ D FRQVWDQW IRUFH SHUKDSV LQ RUGHU WR NHHS D SDUW LQ SODFH ZKLOH WKH RWKHU ÀQJHU PRYHV LQWR D QHZ JULS SRVLWLRQ :KDW PDNHV WKLV D SRWHQWLDOO\ GLIÀFXOW SUREOHP LV WKDW WKH ÀUVW ÀQJHU LV QRW IXOO\ PRELOH ZLWK UHVSHFW WR WKH VHFRQG ÀQJHU ,Q YLHZ RI )LJXUH  QRWLFH WKDW WKH WZR ÀQJHUWLSV PD\ RQO\ PRYH LQGHSHQGHQWO\ LQ D SODQH WKH ZULVW PXVW VZLYHO LQ RUGHU WR PRYH RQH RI WKH ÀQJHUV RXW RI SODQH ,W LV QRW SRVVLEOH WR PRYH RQH ÀQJHU WR DQ DUELWUDU\ SRVLWLRQ ZKLOH NHHSLQJ ERWK WKH SRVLWLRQ DQG RULHQWDWLRQ RI WKH VHFRQG ÀQJHU À[HG ,I WKH SRVLWLRQ UHIHUHQFH RI WKH ÀUVW ÀQJHU FDQQRW EH IXOO\ UHDFKHG ZKLOH PDLQWDLQLQJ WKH SRVH DQG LPSHGDQFH RI WKH VHFRQG ÀQJHU WKHQ RQH RU WKH RWKHU REMHFWLYH

DQG VKRXOG RQO\ EH GLVFDUGHG ZKHQ VLJQLÀFDQW DGYDQWDJHV DUH RIIHUHG E\ D GXDOSULRULW\ ODZ  6LQJOHSULRULW\ FRQWURO ODZ 7KH VLQJOHSULRULW\ ODZ LV ˙ u = J T f ∗ + M J + J˙q˙ + M N Bj q.



7KLV DFKLHYHV D GHVLUHG LPSHGDQFH Λs x ¨s + fs∗ = fs , p1 , ω˙ 1 , p¨2 ) LV D JHQHUDOL]HG DFFHOHUDWLRQ Λs LV ZKHUH x ¨Ts = (¨ WKH SDVVLYH PDQLSXODWRU LQHUWLD SURMHFWHG LQWR WKH FRUUHVSRQG LQJ RSHUDWLRQDO VSDFH FRRUGLQDWHV DQG fs LV WKH JHQHUDOL]HG IRUFH %RWK GHVLUHG LPSHGDQFHV DUH FRPELQHG LQWR D VLQJOH LPSHGDQFH REMHFWLYH ⎛ p ⎛ p ⎞ ⎞ K1 B1 0 0 0 0 fs∗ = ⎝ 0 B1r 0 ⎠ x˙ s + ⎝ 0 K1r 0 ⎠ x ˜s , 0 0 K2p 0 0 B2p )LJ  7KH ELIXUFDWHG PDQLSXODWRU QRW GUDZQ WR VFDOH XVHG LQ WKH VLPXODWLRQV $ 325 LV GHÀQHG DW WKH HQG RI HDFK ´ÀQJHUµ

)LJ  7KH PDQLSXODWLRQ VFHQDULR 7KH WZRÀQJHU PDQLSXODWRU LV LQLWLDOO\ FRQWDFWLQJ WKH REMHFW OLJKWO\ $W WKH VWDUW RI WKH VLPXODWLRQ WKH ULJKW ÀQJHU LPSHGDQFH SRVLWLRQ UHIHUHQFH FKDQJHV VR WKDW WKH ÀQJHU ZLOO PRYH LQ WKH GLUHFWLRQ RI WKH GDVKHG DUURZ 7KH OHIW ÀQJHU LPSHGDQFH UHPDLQV XQFKDQJHG VR WKDW LW ZLOO FRQWLQXH WR DSSO\ D OLJKW LQZDUG IRUFH RQ WKH REMHFW 7KH GXDOSULRULW\ FRQWURO ODZ PXVW PRYH WKH ULJKW ÀQJHU ZLWKRXW FKDQJLQJ WKH LPSHGDQFH RI WKH OHIW ÀQJHU

PXVW EH ´VDFULÀFHGµ ,Q WKLV DSSOLFDWLRQ VFHQDULR LW LV GHVLUDEOH WR VDFULÀFH WKRVH DVSHFWV RI WKH ÀUVW ÀQJHU REMHFWLYH VR WKDW UHJDUGOHVV RI KRZ WKH ÀQJHU LV FRPPDQGHG WR PRYH WKH VHFRQG ÀQJHU LV DVVXUHG RI PDLQWDLQLQJ D FRQVWDQW LPSHGDQFH 7KH HQWLUH PDQLSXODWRU LV LOOXVWUDWHG LQ )LJXUH  7KH WKUHH OLQNV SUR[LPDO WR WKH ELIXUFDWLRQ ZLOO EH UHIHUUHG WR LQ RUGHU SUR[LPDO WR GLVWDO DV WKH ´XSSHU DUPµ WKH ´IRUHDUPµ DQG WKH ´SDOPµ 7KH WZR EUDQFKHV ZLOO EH UHIHUUHG WR DV WKH ´ÀQJHUVµ 7KH OHQJWKV PDVVHV DQG PRPHQWV RI LQHUWLD DUH JLYHQ LQ 7DEOH , 1RWH WKDW WKH ÀQJHU PDVVHV DQG LQHUWLDV DUH ODUJH HQRXJK WR EH RI WKH VDPH RUGHU DV WKH DUP $ 325 LV GHÀQHG DW WKH WLS RI HDFK RI WKH WZR ÀQJHUV % &RPSDULVRQ ZLWK D VLQJOHSULRULW\ FRQWURO ODZ )RU UHIHUHQFH WKH GXDOSULRULW\ FRQWURO ODZ SURSRVHG LQ WKLV SDSHU ZDV FRPSDUHG WR D VLQJOH SULRULW\ FRQWURO ODZ &RPSDUHG ZLWK WKH GXDOSULRULW\ FRQWURO ODZ WKLV LV D VLPSOHU DOWHUQDWLYH

  150

ZKHUH K1p = diag(45) 1HZWRQV SHU 0HWHU DQG B1p = diag(25) 1HZWRQVHFRQGV SHU 0HWHU DUH WKH GHVLUHG WUDQVOD WLRQDO VWLIIQHVV DQG GDPSLQJ RI 325  K1r = diag(10) DQG B1r = diag(5) DUH WKH GHVLUHG DQJXODU VWLIIQHVV DQG GDPSLQJ RI 325  DQG K2p = diag(20) DQG B2p = diag(25) DUH WKH GHVLUHG WUDQVODWLRQDO VWLIIQHVV DQG GDPSLQJ RI 325  7KH FRUUHVSRQGLQJ -DFRELDQ LV ⎛ p ⎞ J1 J = ⎝ J1r ⎠ , J2p

ZKHUH J1p DQG J1r GHVFULEH WKH &DUWHVLDQ WUDQVODWLRQDO DQG DQJXODU YHORFLWLHV UHVSHFWLYHO\ DW 325  DQG J2p LV WKH WUDQVODWLRQDO -DFRELDQ IRU 325   'XDOSULRULW\ FRQWURO ODZ 7KH VLQJOHSULRULW\ FRQWURO ODZ DERYH ZDV FRPSDUHG WR WKH GXDO SULRULW\ FRQWURO ODZ RI (TXDWLRQ  7KH ÀUVW SULRULW\ REMHFWLYH ZDV WR UHDOL]H D WUDQVODWLRQDO DQG DQJXODU GHVLUHG LPSHGDQFH DW 325  ¨1 + f1∗ = f1 . Λ1 x p1 , ω˙ 1 ) LV WKH DFFHOHUDWLRQ LQ SE(3) Λ1 LV WKH FRUUH x ¨1 = (¨ VSRQGLQJ SDVVLYH RSHUDWLRQDO VSDFH LQHUWLD f1∗ LV WKH GHVLUHG GDPSLQJ DQG VWLIIQHVV  p  p   K1 B1 0 0 f1∗ =  + x ˜1 , x ˙ 1 0 K1r 0 B1r

DQG f1 LV WKH FRUUHVSRQGLQJ H[WHUQDOO\ DSSOLHG ORDG 7KH VHFRQG SULRULW\ REMHFWLYH ZDV WR UHDOL]H MXVW D WUDQVODWLRQDO LPSHGDQFH DW 325  ² QRW DQ DQJXODU LPSHGDQFH Λ2 p¨2 + f2∗ = f2 ,

ZKHUH Λ2 LV WKH SDVVLYH PDQLSXODWRU LQHUWLD SURMHFWHG LQWR WKH WUDQVODWLRQDO (XFOLGHDQ FRRUGLQDWHV IRU 325  DQG f2 LV WKH H[WHUQDOO\ DSSOLHG IRUFH DW 325  7KH GHVLUHG 325  GDPSLQJ DQG VWLIIQHVV LV f2∗ = B2p p˙ 2 + K2p p˜2 .



)LJXUH  FRPSDUHV WKH SHUIRUPDQFH RI WKH WZR FRQWURO ODZV ZKHQ WKH ÀUVWSULRULW\ 325 UHIHUHQFH SRVH LV WKH VDPH DV

/LQN XSSHU DUP IRUHDUP SDOP ÀQJHU OLQN

/HQJWK P P P P

Ixx    

0DVV .J .J .J .J

Iyy    

Izz    

7$%/( , '@ $ /LHJRLV ´$XWRPDWLF VXSHUYLVRU\ FRQWURO RI WKH FRQÀJXUDWLRQ DQG EH KDYLRU RI PXOWLERG\ PHFKDQLVPVµ ,((( 7UDQV 6\VW 0DQ &\EHUQHWLFV SS ²  >@ @ / 6FLDYLFFR DQG % 6LFLOLDQR ´$ VROXWLRQ DOJRULWKP WR WKH LQYHUVH NLQH PDWLF SUREOHP IRU UHGXQGDQW PDQLSXODWRUVµ ,((( -RXUQDO RI 5RERWLFV DQG $XWRPDWLRQ YRO  QR   >@ / 6HQWLV DQG 2 .KDWLE ´6\QWKHVLV RI ZKROHERG\ EHKDYLRUV WKURXJK KLHUDUFKLFDO FRQWURO RI EHKDYLRUDO SULPLWLYHVµ ,QWHUQDWLRQDO -RXUQDO RI +XPDQRLG 5RERWLFV  >@ 5 3ODWW 2 %URFN $ + )DJJ ' .DUXSLDK 0 5RVHQVWHLQ - &RHOKR 0 +XEHU - 3LDWHU ' :KHHOHU DQG 5 *UXSHQ ´$ IUDPHZRUN IRU KXPDQRLG FRQWURO DQG LQWHOOLJHQFHµ LQ 3URFHHGLQJV RI WKH  ,((( ,QWHUQDWLRQDO &RQIHUHQFH RQ +XPDQRLG 5RERWV 2FWREHU  >@ 3 &KLDFFKLR 6 &KLDYHULQL / 6FLDYLFFR DQG % 6LFLOLDQR ´&ORVHGORRS LQYHUVH NLQHPDWLFV VFKHPHV IRU FRQVWUDLQHG UHGXQGDQW PDQLSXODWRUV ZLWK WDVN VSDFH DXJPHQWDWLRQ DQG WDVN SULRULW\ VWUDWHJ\µ 7KH ,QWHUQDWLRQDO -RXUQDO RI 5RERWLFV 5HVHDUFK YRO  QR  SS ²  >@ 6 &KLDYHULQL ´6LQJXODULW\UREXVW WDVNSULRULW\ UHGXQGDQF\ UHVROXWLRQ IRU UHDOWLPH NLQHPDWLF FRQWURO RI URERW PDQLSXODWRUVµ ,((( 7UDQVDFWLRQV RQ 5RERWLFV DQG $XWRPDWLRQ YRO  QR  SS ²  >@ * $QWRQHOOL ´6WDELOLW\ DQDO\VLV IRU SULRULWL]HG FORVHGORRS LQYHUVH NLQHPDWLF DOJRULWKPV IRU UHGXQGDQW URERWLF V\VWHPVµ ,((( 7UDQVDFWLRQV RQ 5RERWLFV YRO  QR  SS ²  >@ 0 +XEHU ´$ K\EULG DUFKLWHFWXUH IRU DGDSWLYH URERW FRQWUROµ 3K' GLVVHUWDWLRQ 8 0DVVDFKXVHWWV  >@ 1 0DQVDUG DQG ) &KDXPHWWH ´7DVN VHTXHQFLQJ IRU VHQVRUEDVHG FRQWUROµ ,((( 7UDQVDFWLRQV RQ 5RERWLFV YRO  QR  SS ²  >@ 1 +RJDQ ´,PSHGDQFH FRQWURO  DQ DSSURDFK WR PDQLSXODWLRQ WKHRU\µ -RXUQDO RI G\QDPLF V\VWHPV PHDVXUHPHQW DQG FRQWURO YRO  SS ² 

  152

>@ & 1DWDOH % 6LFLOLDQR DQG / 9LOODQL ´6SDWLDO LPSHGDQFH FRQWURO RI UHGXQGDQW PDQLSXODWRUVµ LQ ,((( ,QW·O &RQI RQ 5RERWLFV DQG $XWRPDWLRQ  SS ² >@ $ $OEX6FKIIHU DQG * +LU]LQJHU ´&DUWHVLDQ LPSHGDQFH FRQWURO WHFK QLTXHV IRU WRUTXH FRQWUROOHG OLJKWZHLJKW URERWVµ LQ ,((( ,QW·O &RQI RQ 5RERWLFV DQG $XWRPDWLRQ  SS ² >@ CD9DED5 1B>5795 #5